Remove background on one file, and a first choice for how
Desktop app
Remove background in Media Studio's One file: add it to a picture or a clip. A green or blue screen gets the colour key, anything else AI (Fast, for people, comes with GlirOS). Make it saves a see-through PNG for a picture, a see-through WebM for a clip (its sound kept), or an MP4 on a colour or a blurred background.
Settings > Tools > Media Studio > "Remove background: try first": Automatic, the colour key, AI or Corridor Key. The chat and One file use it when you just say "remove the background". A picture in the chat with no green screen is now cut out with AI right there.
Corridor Key works alongside: in Remove background, Export for Corridor Key saves the clip's frames and GlirOS's hint in your Videos folder; run Corridor Key on them yourself, then Import its matte. Corridor Key is a separate tool by Corridor Digital, under its own licence. GlirOS doesn't include or run it.
Easier with the keyboard and a screen reader: a "Skip to the message box" link first, Tab stays inside Settings and dialogs (Escape closes them and goes back), the model menu moves with the arrow keys, and a screen reader reads an answer once it is written, not word by word. Uninstall in Windows' Settings > Apps opens GlirOS's dialog in front; a dialog asked for twice shows once, and the media tools' download shows one progress note.
More by keyboard: on Media Studio's timeline a clip moves earlier or later with Ctrl+← and Ctrl+→ (or Earlier and Later in its part), and a screen reader hears its place. The welcome says each step to a screen reader, and its language choice works with the arrow keys. Choices in Settings that look like a row of buttons are one Tab stop and change with the arrow keys too.
Media gallery: mark a part of a video (From/To under the preview) and export just that part into the chat, Exact to the frame or Quick; save a copy into a folder you pick; select several files and send them to Media Studio's timeline in order, each trimmed to its part; and gather pictures and videos from any of your folders into groups, with sub-folders shown as a tree. Your original files are never changed. On the phone, Download says the size first and packs several files into one zip.
Tasks: "Due soon" at the top of the Tasks view lists what is overdue or due within a week, soonest first, from this chat and, under its own heading, from your other chats (a click on a chat's name opens it). Tick them right there; each stays in its own list.
Settings > Remote access shows how quick the connection to your phone or browser is while it is open (the round trip through the relay), with Measure now.
Fixed: sending a part from the media gallery's full-screen viewer to Media Studio left GlirOS not answering clicks. The full screen now closes when the gallery steps aside.
Remote access: the chat list on the PC and on your phone follows what you do on the other within a few seconds: a new chat shows up, a deleted one goes, and an open chat shows the other device's messages as they come. No restart needed.
Media Studio: Record a clip, next to Take a photo, records a short video from your camera (at most a minute, with sound if you like) for One file or the timeline. It stays on this PC, and the camera is off again as soon as you're done.
Media Studio, from your tests: Record a clip no longer plays the microphone while you set up (a level meter shows it instead, so nothing can howl); Ctrl+V with a file open now puts the pasted picture in its place, and says so when there's only text; the Timeline tab can leave the file in One file; a JPG is called JPEG.
Slideshows: add pictures and clips to a chat (or send a group from the Media gallery) and ask "make a slideshow of these". One card lets you choose the pace, your own music file (or Fit to the music), a title, the shape and a slow zoom; Make it puts the MP4 in the chat, and Open on the timeline lets you change anything. Media Studio's timeline also has Slow zoom (In or Out) on any picture.
Media Studio's Looks: the Looks menu in Adjust (One file and a clip on the timeline) has 26 looks with a small preview of each: the ten colour looks, Film (Old film, Cinematic, Dreamy, Grain…), Print & art (Newspaper, Sketch, Pixel art…) and Fun (Heatwave, Thermal, Negative, Mirror, VHS). A look sets the colour and opens the effects it uses as their own panels under Adjust, so you can change any of it; Reset puts the look back, and Save as mine keeps your own. Or say it in the chat: "make it look like an old film".
Web search can use your own SearXNG server (a free search engine you run yourself, on a home server, a NAS or a VPS): Settings > Web search > SearXNG, put in its address and press Test and save. Searches go straight from your PC to your server; GlirOS doesn't run or provide one. A server behind a password: open "This server asks for a name and password" under the address; they are kept in GlirOS's key store, never in the settings file.
A research report's Save as PDF now saves the PDF straight away: pick where in Windows' Save dialog, no print dialog. The PDF has bookmarks for its sections and is tagged for screen readers; Open on the notice that follows opens it. Uninstall started from Windows' Settings > Apps on a first start shows the uninstall dialog, not the welcome; a GlirOS of another Windows account says so.
Replace text in a video now follows a text that moves with the camera or the sign; a cut, something passing in front, a zoom or changing light is said in words, and nothing is made. With a clip open in Media Studio, the chat does it too (replace "SALE" with "SALG"), and pictures and clips on the timeline have Replace text under Add a part, the result going in their place.
Face swap has two new choices in its AI steps: Stronger likeness on Swap (the swap runs twice: a little more like the person and less of the face it replaces) and Detail under Restore (the swap made at up to four times the size, so a big face comes out sharp while keeping the most likeness; a small face stays as before). Both take longer. Opened again, a chat keeps the face to put in with its picture and name, also after a New chat. Its faces are found again too.
Import chats from Claude and Gemini too: Settings > Chat history > Import chats takes Claude's export (Settings > Privacy > Export data) and Gemini's Google Takeout (My Activity, Gemini Apps, as JSON) beside ChatGPT's. The file is read on your device only; importing it again adds nothing.
Workspaces in Media Studio's timeline: the Workspace menu (top right) has Edit, Colour, Key, Sound, Text and Export (keys 1–6) and your own (Save, Update, Manage). A workspace keeps which panels stay in view, never a clip's values: in Colour, Colour and Looks stay for every clip you choose, quiet with "+ Apply to this clip" where they aren't on it yet. Each chat remembers its workspace.
Media Studio's timeline: a chosen clip or layer shows its name once with what is on it ("On this clip: Colour · Looks", each with × to take it off), its settings, and each part as a panel you can fold, reset to neutral, remove or hide. Hiding a panel never takes the part off. A rail of parts on the side (Colour, Looks, Crop, Key) adds a part with one tap, shows a dot when it is on the clip, and brings a hidden panel back.
On the timeline, Ctrl+click (or Shift+click) more clips to change them together: a slider, Remove or a look goes to all of them. Copy parts / Paste parts (top of the side) put one clip's Colour, Looks, Crop and Key on others. Drag a panel's header (or Alt+↑/↓) to reorder, and P pins.
Looks on the timeline's layers too (a logo or a picture-in-picture: its own colour and look, its see-through parts kept), and a look asked for in the chat ("make it look like an old film") goes on the timeline's chosen clip or layer when the timeline is open.
The welcome screen's suggestions fit the model you use: short, clear asks for a small model (five tips for better sleep, an explanation in three sentences, a short email), what to see on a day trip to Lillehammer for a middle-sized one, and the full Lofoten trip plan only for a big model or a cloud model.
Your own SearXNG server behind a name and password: when its address is plain http and not this PC or your home network, Settings > Web search says the name and password are sent unencrypted. If the server sends the search on to another address, GlirOS no longer sends the name and password there; it tells you to set the address the search ends up at.
Media Studio's panels: the eye now only ever means "seen in the preview". On the timeline a panel is hidden with a tap on its lit part in the rail (or P), in Edit too, and the part stays. One file's parts are panels like the timeline's: fold, the name with its preview eye, and the bin to remove. Sliders keep their value on the same line, Colour starts with the Clip panel folded, and the top button says Remove clip or Remove layer.
Media Studio: One file's colour part is now called Colour, the same word as on the timeline. The chosen clip's name keeps its room beside Remove clip, and One file's buttons at the bottom no longer squeeze their words onto two lines.
Tools you set up and run yourself are marked Advanced and no longer stand beside the built-in choices. Settings > Web search has four choices, and your own SearXNG server sits in a fold under them, "Your own search server". Remove background shows Corridor Key only when it is found on this PC or chosen in Settings. GlirOS works fully without either.
Media Studio's One file now has the timeline's side: "On this file" with a chip per part, the parts as panels, and the rail on the outer edge (Look · Shape and time · AI · Out; a part this file can't take is greyed and says why). The Workspace menu is on One file's bar too, with the same six workspaces, keys 1–6 and your own, shared with the timeline: Colour pins Colour and Looks, with "+ Add to this file".
Research reports: a long section folds after its first paragraph (printing opens it), a citation shows its source on hover, focus or a tap, and a Figures tab in the appendix lists every chart, map and set of steps. Export now puts the report's PDF first in its zip, beside the page, the text and the pictures. The plan has a question for each part of what you asked (Covers: shows which), adds no year you didn't give, and the Summary says what was partly found.
Replace text in a video now follows a sign that comes closer or turns a little (up to 1.25 times its size and 12 degrees): the new words grow and turn with it. One that turns or zooms further is refused with that reason, and nothing is made.
Replace text has Translate beside the new words: it fills them with the found text in this app's language, translated by the chat's model (a cloud model asks first; only the text is sent, not the picture). You can change it before you replace.
Settings > About: when the update check can't finish, it now says why in your language (the update file isn't on GlirOS's website yet, the website couldn't be reached, or Block all network traffic is on).
Connectors work with servers on MCP's newest version (July 2026) as well as the older ones: GlirOS asks a server which version it speaks first, then talks to it that way. A server that speaks only a later version says so in Settings > Connectors.
Research reports can show a picture from a source: only when the source's own page shows a free licence for it (public domain, CC0, CC BY or CC BY-SA), with its full credit, and saved with the report, so opening it later loads nothing from the site. Any other source picture is shown as a card that links to the page, with no picture.
Connectors can sign in to services: when a service needs it, its card in Settings > Connectors says so, and Sign in opens the service's own page in your browser. GlirOS never sees your password; the access it gives is kept in Windows' key store, renewed by itself, and Sign out (or removing the connector) gives it back.
Research can add a few pictures made by AI: choose A few (up to 3) in the brief. They come from your picture model, only for sections about something that can be seen, and each says "Made by AI" with the model's name. Off unless you choose it.
Small models on your device get a lighter prompt: the tools are named, and only the calls of the tools a message is about go with it (at most 3 for a small model), so the right tool comes first far more often. Gemma 4 E2B now counts as the 2B model it runs as, and GlirOS leads it through big jobs step by step. Under each answer, How this answer was made shows what the model read, part by part.
Fixed: the globe (web search) pressed in a new chat before its first message was lost; that message now searches the web, and the globe stays on for the chat.
Skills: small functions the chat fetches the first time a question needs one, for exact answers a model gets wrong. The first: Exact maths, Calendar maths (working days, holidays, sun and moon), the Cube solver, Text and lines (counts, sorting, compare) and Everyday calculators (loans, savings, BMI; not advice). The chat asks first: "Fetch the Exact maths skill from tools.gliros.com? Nothing about your question is sent." The answer shows "Worked out by the … skill". Settings > Tools lists the skills fetched, with Remove.
One file: Looks has a panel of its own beside Colour (as on the timeline), a greyed part says why ("A picture has no sound."), the side scrolls on its own in full screen, and the preview's background is no longer black.
Settings > Pictures > Advanced: Ask before sending now says which companies you have said yes to (Ask again is greyed when there are none), and refreshing the picture model list has its own row that says what it does.
Cloud models: Claude now reuses the unchanged start of each request from Anthropic's cache (much cheaper on long chats); a message a switched-on tool answers ("convert 5 miles to km", "how many days to Christmas Eve") is no longer also searched on the web; and "17.5 % of 2 348.60" is read with its space between thousands.
When a switched-on tool should answer but doesn't (the model doesn't call it, the call fails or finds nothing), GlirOS searches the web after all and the model answers again from the results, with the sources under it. A small model's answer that says it's only a simplified version gets a bigger model offered under it.
Code: a long code block has its buttons (preview, side panel, copy) at its end too, and in the side panel's list of files one click on a file opens it and shows its versions.
Research: when pictures made by AI were chosen and none was made, the report and the research card say why (for example, no section is about something that can be seen), and the pictures no longer get made-up writing in them. A page that never names the subject is left out; Links lists every page read and why. Export as zip asks where to save it. A page too long for the model is read again with less of it, and a second report of the same day and title gets its own folder. A chart of years is left out.
Connectors with a cloud model: each call now goes with the service's own description of its fields (which are optional, which take a list), so a service like Notion accepts it, and an empty optional field is left out before it is sent. A call the model gets wrong is answered with the service's calls, so the model tries again instead of ending with an empty answer. The tools stay the same through the whole answer, so Claude can reuse its cache.
A cloud model can now take up to 8 tool steps in one answer (a model on this device still 3). When the limit stops a call, a line under the answer says so: say "go on" to let it continue.
How this answer was made > What the model read: the tools a cloud model can call (sent as functions) have a line of their own, so the other lines show their real size.
Fixed: in the side panel, One file's View menu opened off the screen; it now opens just under the View button.
Media gallery: Make a slideshow on a group's header or the Select bar puts its pictures and clips in the chat with the slideshow card (pace, your music, the title from the group's name, the shape), then Make it.
Skills: Everyday calculators work out a pregnancy's due date, its weeks and trimester; general information, not medical advice, and its calls aren't written to the chat log.
Connectors with a cloud model: fewer, smaller tool descriptions go with each request (well under half of before for a big service like Notion), fields with free names (a Notion page's title) are described as such, and "go on" keeps the calls the job needs.
The cloud token limit for one answer counts tokens the company reads from its cache at a tenth, as they cost.
Connectors: services that describe their fields with references, choices of shapes or merged parts (common in servers written with Pydantic or Zod) are now passed to a cloud model in a form every company takes, with the choices said in words; a number or list the model writes as text is fitted before it is sent.
Media gallery: Send to > The chat puts a picture or clip into the message you are writing, so the model can look at it, and from there One file can pick it.
Media gallery: Send to Media Studio with one file opens it in One file — a picture stays a picture — while several files or a part go to the timeline as before.
Asking about the picture in the side panel ("can you see the picture in the side panel?") gives the model what the open Media gallery shows, the file open in it included.
The calendar answers "is it Halloween today?" and a short follow-up like "is it today?" after "when is Halloween?", and gives its answer to the model as the one to use.
With the Calendar & seasons tool switched on, a question about a day ("when is Easter?", "is it Halloween today?") now gets the calendar's answer from the same rules too, instead of waiting for the model to call the tool.
Special days: the chat knows every day of the calendar's rules that falls today, the world's too (Women's Day, St Patrick's Day, Halloween…), and mentions it once with a short greeting where it fits.
The calculator no longer works out a range, a score or a fraction in an ordinary request ("plan a 3-4 day trip", "the meeting is 2-3pm", "1/2 cup sugar"). It runs by itself when you ask for a sum ("what is…", "regn ut…") or the message is mostly the sum, and "1500 kr pluss 25 % mva" is worked out.
Everyday requests run by GlirOS itself, even with a small model: "remember to bring out the garbage every thursday" (it asks what time), "every morning at 7 tell me the weather", "will it rain tomorrow in Bergen?", "what is 5 feet 10 in cm?", "the exchange rate for pounds today", "pick one of pizza, taco or sushi", "how many words are in this text: …".
Settings > Security says who can read what GlirOS keeps on this PC, and how to protect it (a password or PIN, locking the screen, BitLocker).
My calendar answers "what do I have on Friday?", "what's on my calendar next week?" and "am I free on Thursday afternoon?" by itself, even with a small model. Adding and moving appointments still goes through the model, which asks you first.
With many tools switched on, a cloud model now gets only the tools the message is about (and the one open in the side panel), not all of them: each request is much smaller and cheaper. Connectors keep their own limit.
Reminders from plain words: "remember to bring out the garbage every Thursday" asks what time and sets it; "change the garbage reminder to 19:00" moves it, and each reminder in Timers has a ✎ to change its time, repeat or text. My calendar now shows your reminders on their days (🔔) and names them when you ask "what do I have on Thursday?".
Connectors: GitHub's connector works again after GlirOS restarts, a program on this PC started with npx or uvx now starts on Windows, and when such a program stops, its own error line is shown. A small model gets the read call with a list call, is told which folder a program on this PC works in, and makes a call that only reads without asking you first.
When the model offers to do something ("Would you like me to read the file?") and you say yes, GlirOS tells it to do it now with the tool, never from memory. The Everyday calculators skill is found for questions like "when is the baby due if the last period started 1 March?".
Reading a web page checks the address it really connected to, so a site name that suddenly points into your own network isn't read.
Speak the subtitles starts with the voice you chose for the language in Settings > Voice, the same one read-aloud uses. The dubbed file shows its real size, and the chat's line says it in words.
New built-in tool, Voices: choose the voice for each language (it is used everywhere GlirOS reads aloud), hear a sample, and turn a text into a sound file in the chat's Files. Timers reads reminders aloud with your chosen voice. "Read this aloud: …" and "use Kokoro for English" are done by GlirOS itself; Save asks where; the files carry no encoder tag. A file GlirOS already made for the message isn't made again.
Lighter instructions for the model: the rules for plans, to-do lists and media editing come only when a message needs them, and what GlirOS adds to a message is kept within a smaller model's room, so more of the conversation fits.
With a small model, GlirOS now works out a due date, a BMI or a loan itself when you ask in plain words ("when is the baby due if the last period started 1 March?"), with the Everyday calculators skill, and the model tells you the result.
Projects with a folder on your PC: a web page opened from the folder now previews with its own CSS and JavaScript from that folder, and Open as a code project shows the folder's files (no longer files left from an older project of the same name). When you ask to change something in a small project, the model gets its code files whole. Pages that save to localStorage (a game's high score) now run in the preview.
Mix versions in the preview: in Files and versions, the pin on a version makes the preview use it (for example the script's v2 with the newest page and the first version of the styles). The pins stay while you open other files, the code project opens with them, and the x goes back to the newest. Nothing is deleted.
When the model answers with just the changed lines of a file, GlirOS puts them into the whole file itself, so the rest of the code is never lost: a card under the answer shows the change, Apply puts it in (in the project folder, or as a new version), and Test it in the preview runs it. A new script or stylesheet gets a card to link it from the page, which keeps its other links. The Files panel now says where each file is: only in this chat, the same as in the project folder on your PC, or different from it.
Files and versions can rename a file (its versions go with it, and a project folder's file of that name too), and so can the code project. A file deleted in Files and versions now leaves the open code project as well, and the file list fills the panel when nothing is shown below it.
Maths a model writes in LaTeX, like $1234 \times 56$, now shows and is read aloud as plain maths: 1234 × 56. Prices with a dollar sign are left as they are.
The sidebar shows project names in full (the number of chats is in the tooltip), a pulse where an answer is being written, and a dot on a chat, and on its project even when it is folded, when an answer finished while you were away, until you look at it.
Code in a project folder with a small model: a small folder's code now always goes with a message about changing it, and an answer that writes its lines with line numbers ("73 ctx.fillStyle …") is put into the right lines of the file. A change applied in the folder offers Test it in the preview. Linking a folder starts in Documents, and the chat's notes about what you did are in your language.
Code in a project folder: a page in the preview keeps what it saves (a game's high score) when you reload the preview; asking for a new script in a project no longer brings GlirOS's ready-made examples; a whole page a model writes without saying its language is taken as the page, with Use this as index.html.
An answer's code block that leaves parts out ("// ... existing code") is put into the file part by part, or not kept at all: it never becomes a half file the preview would run. A change card for a page, script or stylesheet has Test it in the preview on it, and when one card is applied, a card with the same change says so too.
The desktop app keeps what its pages remember over a restart (a version you pinned for the preview, the chats with new answers), and the Code workspace's Live preview keeps a page's saved data over ⟳. A file you name in a chat with a files connector is read for the model straight away, and a question about a connector's own things (its issues) isn't searched on the web first. While a model writes a change or a tool call, a line shows it is coming.
On a graphics card with 6 GB or less, the automatic context size now stops at 16 384 tokens instead of 32 768 (8 192 for a model whose conversation memory would take more than a quarter of the card): the small models answer faster there, with the same quality in our tests. Long chats are summarised as before, and you can still set a bigger size by hand in Settings › Models › Advanced.
Web version
Easier with the keyboard and a screen reader: a "Skip to the message box" link first, Tab stays inside Settings and dialogs (Escape closes them and goes back), the model menu moves with the arrow keys, and a screen reader reads an answer once it is written, not word by word. A dialog asked for twice shows once, and the media tools' download shows one progress note.
The GlirOS Helper no longer runs out of memory on longer questions (for example a question about GlirOS itself, which brings the help pages along). It downloads its model once more. If a model in the browser still runs short of memory, GlirOS tries again with less and says so in plain words.
Remove background in the web version: "Fast — for people" now runs in your browser, so a person can be cut out of a picture or a clip without a green screen. A picture comes back as a PNG with a see-through background; a clip is put on a colour, Blur or a screen as an MP4. Nothing is uploaded. A see-through video, "Best — any subject" and Corridor Key stay in the desktop app.
More by keyboard: on Media Studio's timeline a clip moves earlier or later with Ctrl+← and Ctrl+→ (or Earlier and Later in its part), and a screen reader hears its place. The welcome says each step to a screen reader, and its language choice works with the arrow keys. Choices in Settings that look like a row of buttons are one Tab stop and change with the arrow keys too.
Media gallery: export pictures and videos into the chat (whole files in the web version) or save copies into a folder you pick; select several files and send them to Media Studio's timeline in order, each trimmed to the part you marked; and gather files from any of your folders into groups, with sub-folders shown as a tree. Your original files are never changed.
Tasks: "Due soon" at the top of the Tasks view lists what is overdue or due within a week, soonest first, from this chat and, under its own heading, from your other chats (a click on a chat's name opens it). Tick them right there; each stays in its own list.
My PC: the model button shows whether your PC can be reached: a grey dot and "offline" when it is off or asleep, green again as soon as it answers.
Settings > Remote access shows which GlirOS account this browser is signed in to, with Sign out, or Sign in when it isn't.
Media Studio's Looks: the Looks menu in Adjust (One file and a clip on the timeline) has 26 looks with a small preview of each: the ten colour looks, Film (Old film, Cinematic, Dreamy, Grain…), Print & art (Newspaper, Sketch, Pixel art…) and Fun (Heatwave, Thermal, Negative, Mirror, VHS). A look sets the colour and opens the effects it uses as their own panels under Adjust, so you can change any of it; Reset puts the look back, and Save as mine keeps your own. Or say it in the chat: "make it look like an old film".
Import chats from Claude and Gemini too: Settings > Chat history > Import chats takes Claude's export (Settings > Privacy > Export data) and Gemini's Google Takeout (My Activity, Gemini Apps, as JSON) beside ChatGPT's. The file is read on your device only; importing it again adds nothing.
Workspaces in Media Studio's timeline: the Workspace menu (top right) has Edit, Colour, Key, Sound, Text and Export (keys 1–6) and your own (Save, Update, Manage). A workspace keeps which panels stay in view, never a clip's values: in Colour, Colour and Looks stay for every clip you choose, quiet with "+ Apply to this clip" where they aren't on it yet. Each chat remembers its workspace.
On the timeline, Ctrl+click (or Shift+click) more clips to change them together: a slider, Remove or a look goes to all of them. Copy parts / Paste parts (top of the side) put one clip's Colour, Looks, Crop and Key on others. Drag a panel's header (or Alt+↑/↓) to reorder, and P pins.
Media Studio's timeline: a chosen clip or layer shows its name once with what is on it ("On this clip: Colour · Looks", each with × to take it off), its settings, and each part as a panel you can fold, reset to neutral, remove or hide. Hiding a panel never takes the part off. A rail of parts on the side (Colour, Looks, Crop, Key) adds a part with one tap, shows a dot when it is on the clip, and brings a hidden panel back.
Looks on the timeline's layers too (a logo or a picture-in-picture: its own colour and look, its see-through parts kept), and a look asked for in the chat ("make it look like an old film") goes on the timeline's chosen clip or layer when the timeline is open.
The welcome screen's suggestions fit the model you use: short, clear asks for a small model (five tips for better sleep, an explanation in three sentences, a short email), what to see on a day trip to Lillehammer for a middle-sized one, and the full Lofoten trip plan only for a big model or a cloud model.
Media Studio's panels: the eye now only ever means "seen in the preview". On the timeline a panel is hidden with a tap on its lit part in the rail (or P), in Edit too, and the part stays. One file's parts are panels like the timeline's: fold, the name with its preview eye, and the bin to remove. Sliders keep their value on the same line, Colour starts with the Clip panel folded, and the top button says Remove clip or Remove layer.
Media Studio: One file's colour part is now called Colour, the same word as on the timeline. The chosen clip's name keeps its room beside Remove clip, and One file's buttons at the bottom no longer squeeze their words onto two lines.
Media Studio's One file now has the timeline's side: "On this file" with a chip per part, the parts as panels, and the rail on the outer edge (Look · Shape and time · AI · Out; a part this file can't take is greyed and says why). The Workspace menu is on One file's bar too, with the same six workspaces, keys 1–6 and your own, shared with the timeline: Colour pins Colour and Looks, with "+ Add to this file".
Research reports: a long section folds after its first paragraph (printing opens it), a citation shows its source on hover, focus or a tap, and a Figures tab in the appendix lists every chart, map and set of steps. In the browser, Export's zip has no PDF: Save as PDF makes it. The plan has a question for each part of what you asked (Covers: shows which), adds no year you didn't give, and the Summary says what was partly found. A page that never names the subject is left out; Links lists every page read and why.
Replace text has Translate beside the new words: it fills them with the found text in this app's language, translated by the chat's model (a cloud model asks first; only the text is sent, not the picture). You can change it before you replace.
Follow the person in the web version too: a crop to 9:16, 1:1 or 4:5 keeps the main person in the picture as they move (Crop in One file or on a timeline clip, or "crop to 9:16 and follow the person" in the chat). The small face finder comes with the web version and runs in your browser: nothing is sent anywhere.
Research reports can show a picture from a source: only when the source's own page shows a free licence for it (public domain, CC0, CC BY or CC BY-SA), with its full credit, and saved with the report, so opening it later loads nothing from the site. Any other source picture is shown as a card that links to the page, with no picture. A page too long for the model is read again with less of it, and a second report of the same day and title gets its own folder. A chart of years is left out.
Models in your browser get a lighter prompt: the tools are named, and only the calls of the tools a message is about go with it, so the right tool comes first far more often. Under each answer, How this answer was made shows what the model read, part by part.
Fixed: the globe (web search) pressed in a new chat before its first message was lost; that message now searches the web, and the globe stays on for the chat.
Skills: small functions the chat fetches the first time a question needs one, for exact answers a model gets wrong. The first: Exact maths, Calendar maths (working days, holidays, sun and moon), the Cube solver, Text and lines (counts, sorting, compare) and Everyday calculators (loans, savings, BMI; not advice). The chat asks first: "Fetch the Exact maths skill from tools.gliros.com? Nothing about your question is sent." The answer shows "Worked out by the … skill". Settings > Tools lists the skills fetched, with Remove; the web version keeps them in this browser.
One file: Looks has a panel of its own beside Colour (as on the timeline), a greyed part says why ("A picture has no sound."), the side scrolls on its own in full screen, and the preview's background is no longer black.
Settings > Pictures > Advanced: Ask before sending now says which companies you have said yes to (Ask again is greyed when there are none), and refreshing the picture model list has its own row that says what it does.
Cloud models: Claude now reuses the unchanged start of each request from Anthropic's cache (much cheaper on long chats); a message a switched-on tool answers ("convert 5 miles to km", "how many days to Christmas Eve") is no longer also searched on the web; and "17.5 % of 2 348.60" is read with its space between thousands.
When a switched-on tool should answer but doesn't (the model doesn't call it, the call fails or finds nothing), GlirOS searches the web after all and the model answers again from the results, with the sources under it. A small model's answer that says it's only a simplified version gets a bigger model offered under it.
Code: a long code block has its buttons (preview, side panel, copy) at its end too, and in the side panel's list of files one click on a file opens it and shows its versions.
A cloud model can now take up to 8 tool steps in one answer (a model on this device still 3). When the limit stops a call, a line under the answer says so: say "go on" to let it continue.
How this answer was made > What the model read: the tools a cloud model can call (sent as functions) have a line of their own, so the other lines show their real size.
Fixed: in the side panel, One file's View menu opened off the screen; it now opens just under the View button.
Media gallery: Make a slideshow on a group's header or the Select bar puts its pictures and clips in the chat with the slideshow card (pace, your music, the title from the group's name, the shape), then Make it.
Skills: Everyday calculators work out a pregnancy's due date, its weeks and trimester; general information, not medical advice, and its calls aren't written to the chat log.
The cloud token limit for one answer counts tokens the company reads from its cache at a tenth, as they cost.
Media gallery: Send to > The chat puts a picture or clip into the message you are writing, so the model can look at it, and from there One file can pick it.
Media gallery: Send to Media Studio with one file opens it in One file — a picture stays a picture — while several files or a part go to the timeline as before.
Asking about the picture in the side panel ("can you see the picture in the side panel?") gives the model what the open Media gallery shows, the file open in it included.
The calendar answers "is it Halloween today?" and a short follow-up like "is it today?" after "when is Halloween?", and gives its answer to the model as the one to use.
With the Calendar & seasons tool switched on, a question about a day ("when is Easter?", "is it Halloween today?") now gets the calendar's answer from the same rules too, instead of waiting for the model to call the tool.
Special days: the chat knows every day of the calendar's rules that falls today, the world's too (Women's Day, St Patrick's Day, Halloween…), and mentions it once with a short greeting where it fits.
The calculator no longer works out a range, a score or a fraction in an ordinary request ("plan a 3-4 day trip", "the meeting is 2-3pm", "1/2 cup sugar"). It runs by itself when you ask for a sum ("what is…", "regn ut…") or the message is mostly the sum, and "1500 kr pluss 25 % mva" is worked out.
Everyday requests run by GlirOS itself, even with a small model: "remember to bring out the garbage every thursday" (it asks what time), "every morning at 7 tell me the weather", "will it rain tomorrow in Bergen?", "what is 5 feet 10 in cm?", "the exchange rate for pounds today", "pick one of pizza, taco or sushi", "how many words are in this text: …".
My PC on a signed-out GlirOS account: your PC's models stay in the list, greyed with "sign in to use", and Sign out takes them off this browser.
The first time you add a key, GlirOS asks "Is this your own computer?": on a shared or public one the key is kept only until you close the tab. Settings > Security has Forget this browser (keys, chats and settings gone; downloaded models stay).
Opening GlirOS no longer asks Hugging Face for model sizes; the sizes come with GlirOS, and Hugging Face is contacted only when you download a model.
My calendar answers "what do I have on Friday?", "what's on my calendar next week?" and "am I free on Thursday afternoon?" by itself, even with a small model. Adding and moving appointments still goes through the model, which asks you first.
With many tools switched on, a cloud model now gets only the tools the message is about (and the one open in the side panel), not all of them: each request is much smaller and cheaper. Connectors keep their own limit.
Reminders from plain words: "remember to bring out the garbage every Thursday" asks what time and sets it; "change the garbage reminder to 19:00" moves it, and each reminder in Timers has a ✎ to change its time, repeat or text. My calendar now shows your reminders on their days (🔔) and names them when you ask "what do I have on Thursday?".
When the model offers to do something ("Would you like me to read the file?") and you say yes, GlirOS tells it to do it now with the tool, never from memory. The Everyday calculators skill is found for questions like "when is the baby due if the last period started 1 March?".
Safer model downloads in the browser: every model is fetched at a fixed version, so a model's files can't be swapped later. A model you already downloaded is fetched once more the next time you use it.
Reminders in plain words, answered in the language you wrote in: "remind me to call mom every day at five" gets "Reminder set …" even with another language chosen. When GlirOS asks what time and you answer "18", it sets the reminder and says so itself. A task for later, like "give me a recap of the news every day at two o'clock", isn't searched on the web now; the search runs when the task does.
The note to check official sources for tax and money questions no longer appears under ordinary Norwegian messages ("dette" means "this", not debt), or under questions about a nav bar or the benefits of something.
Skills: a question that names a calculator ("my BMI at …", "a loan of …") gets that calculator, a power written in words ("2 to the power of 100") gets Exact maths, and a calculator's caution line is under its card. With the globe on, a question a switched-on tool answers ("5 miles in km") is answered by the tool, not searched as well; asking to search still searches. A message in another language than the one before is answered in its own language, even with only a word or two that tell; and when a tool GlirOS ran answers, a model on your device answers with its result.
Skills: a question a skill works out ("15 % of the working days left this year") isn't searched on the web by itself; How this answer was made names the skills that came with the message; and when Block all is on or the chat is offline, a line says why a skill that isn't on the device wasn't fetched. A due date, BMI or loan in your question is worked out by GlirOS itself for every model on this device, and a refused try isn't shown as a card. Skills you have are ready right after a restart.
Web search: with the globe on, a question is searched even when a tool or skill shares a word with it (only a tool GlirOS runs itself, like a unit conversion, goes first); a small model's own web search is run by GlirOS, with Sources; Gemma's raw call marks no longer show; the search's line is in your language; How this answer was made names the tools that came with the message; and one Open button per tool.
Remove background and Replace text: a pale sky or sea is no longer taken for a blue screen, and "with AI" in the message always means AI; the green-screen card says which question is still open, suggests plain black behind, and shows no stray words; "replace the text STOP with GLIROS" works on a clip added to the chat, finds text on a sign over a busy background, and keeps the sign's board; Media Studio's plan line and the slideshow card speak your language. A clip attached with "replace the text …" is the one used, and a sign in capitals is translated as a sentence.
When a small model's own call to a tool GlirOS already ran fails ("how many km is 26.2 miles?"), the model gets GlirOS's result to answer with; a place it looks up keeps your town ("Karl Johanns gate 22, Oslo"); a stray "[/Web results]" no longer shows; How this answer was made and the buttons under tool calls use the tools' names ("Open Map"); and a question with no numbers only brings a skill when it uses that skill's own words.
Slideshows: a tall (9:16) or square slideshow fills the frame and keeps what stands out in each picture; clips and pictures keep the order you picked; the plan line names the shape; "make the pictures 5 seconds" changes the slideshow on the timeline, which keeps each picture's name; "all these" with files attached takes them and the chat's earlier ones; a slideshow GlirOS made isn't counted as a clip; and a message that can't take all your pictures (8 at a time) says how many were left out.
A music file sent with pictures isn't written down by itself; Stop before anything was sent puts your message back in the box and says so; a later round of an answer starts a new paragraph; and Cube solver 1.0.2: "solve this Rubik's cube scramble: R U R' …" is solved by GlirOS itself, also with a small model, from the moves in your message and checked; a model that calls it with the scramble gets the solution (Everyday calculators 1.0.6 also takes a model's own values first); a new scramble is never shown as a solution.
Media Studio's Looks: a look's slider keeps focus, so the arrow keys move it step by step, and the side stays where it was scrolled (in One file and on the timeline); the look list opens fully in view when its section sits low; and Escape on the list closes the list without making Media Studio smaller. Workspaces: moving a slider with several clips chosen changes only that value on the others; Alt+↓ works right after a fold, Ctrl+Z right after a chip's ×. A workspace just saved is ticked in the menu.
Tools
Voices 1.1: make the sound file as WAV, MP3, M4A or Opus (the desktop app, once Media Studio's engine is installed), then Save it where you choose or send it to Sound Studio or Media Studio. Open a subtitle file (.srt or .vtt) to have it spoken with timing, or a .txt to fill the text box. A voice that can only read aloud now says so, and which voice makes the files. A text longer than 2000 characters is kept as parts (Part 1 of 3, Next part), and Clear starts over.
The Translator and Voices work together: Translate first… in Voices sends the text to the Translator, and Send to Voices in the Translator brings the translation back, ready to be made into a sound file. If one of them is switched off, the other says so. Voices asks and answers in the language you wrote in and keeps a voice's name as it is; the text box and a subtitle file find their own language.
A fast cut-out for people, and Paste or Take a photo in Media Studio
Desktop app
Remove the background without an add-on: under the key's AI chip, choose "Fast — for people". A small model that comes with GlirOS finds the person in a picture or a video, on the processor, in a few milliseconds a frame. For objects, or finer hair, choose "Best — any subject" (the background remover add-on). "Remove the background" in the chat uses Fast when the add-on isn't installed.
Paste and Take a photo in Media Studio: next to "From this device", paste a picture you copied (or press Ctrl+V in the panel), or take one photo with the camera: for One file, the face swap's face, a new target, and the timeline's clips, layers and watermark. The photo stays on your device; the camera turns off when the view closes. Not in the chat.
Research: a Norwegian report reads Norwegian pages before Danish and Swedish ones, and a note a Danish page gave in Danish is written again in Norwegian; a price in US dollars in a report for another country is set aside under "What we could not confirm"; the same finding is no longer listed twice; NDLA counts as a reference.
Web version
Paste and Take a photo in Media Studio: next to "From this device", paste a picture you copied (or press Ctrl+V in the panel), or take one photo with the camera: for One file, a new target, and the timeline's clips, layers and watermark. The photo stays on your device; the camera turns off when the view closes. Not in the chat.
Instructions and memory for projects and chats: in a project's ⋯ menu, write what every chat in it should know or do ("answer in Norwegian; the house is from 1978") and keep facts for the project; a chat's own ⋯ menu does the same for that chat only. They go with every message, the chat's own first. "Remember that …" in a project keeps it for the project (with "Only this chat" on the notice), and outside a project for the chat when Memory across chats is off. A smaller model gets a short version, made once, with every number and date kept. Never in an incognito chat.
Media Studio's colour key has new backgrounds behind the cut-out: Blur (the clip's own picture, blurred, with a strength slider), Green screen and Blue screen (pure colours, so the result can be keyed again in another program), beside a colour.
Research: a research waiting for its plan to be approved, or one you opened again from an old research chat, no longer stops you from starting a new one; deleting a research chat takes its research with it. A report with its moving section headers now opens in the side panel even when it is large. The sources list shows the numbers the text cites, a quote cut mid-sentence runs on to the sentence's end, and the stronger of two agreeing sources is cited.
Settings > Tools: switching a tool on or off keeps the list where it was, so the next tool is right there.
Web version
Instructions and memory for projects and chats: in a project's ⋯ menu, write what every chat in it should know or do ("answer in Norwegian; the house is from 1978") and keep facts for the project; a chat's own ⋯ menu does the same for that chat only. They go with every message, the chat's own first. "Remember that …" in a project keeps it for the project (with "Only this chat" on the notice), and outside a project for the chat when Memory across chats is off. A smaller model gets a short version, made once, with every number and date kept. Never in an incognito chat.
Media Studio's colour key has new backgrounds behind the cut-out: Blur (the clip's own picture, blurred, with a strength slider), Green screen and Blue screen (pure colours, so the result can be keyed again in another program), beside a colour.
Settings > Tools: switching a tool on or off keeps the list where it was, so the next tool is right there.
The desktop app now tells you, once, when a newer GlirOS is out: a quiet notice with what's new and a link to the download page. Once a day it reads a small signed file on GlirOS's website; nothing about you is sent, and nothing is downloaded or installed by itself. Settings > About > Updates has Check now and a switch to turn it off.
My calendar from the chat: a time the model writes under another name ("time" for the start) is no longer lost, so "Wednesday at 11" is at 11 and not all day; "Wednesday next week" is worked out as a date for the model; and an appointment goes in My calendar, not in a to-do list.
Research and web questions: when most searches fail because the free search is busy, GlirOS now says so and that a Brave or Tavily key would help. With a Brave key, each result now brings more text from its page, and a question in Norwegian, German, Spanish, French or Italian asks Brave for pages in that language and country.
Web search: Serper is a new optional choice in Settings > Web search, with your own key.
Delete my data and Uninstall: Settings > Security > Delete my data removes your chats, settings, files and keys from this PC and starts GlirOS again empty (the models are a separate choice). Settings > About > Uninstall GlirOS, or Windows Settings > Apps, removes the app; your data and models are kept unless you tick them. Each says what goes, with sizes, and what was removed at the end.
Research: a sentence citing a page in another language is no longer moved to "could not confirm" (a Norwegian report from English pages lost whole sections), each source's answer is noted in the report's language, the plan looks for an official or reference source first, and a Tavily search for a question in Norwegian, German, Spanish, French or Italian favours that country.
Tools
My calendar 1.2.2: an appointment already in the calendar isn't added twice. With a new note or another time, it asks you whether to update or move it, and only your click decides.
Settings > Models > Advanced > Add a model from a link: paste a model's Hugging Face link, and GlirOS shows its licence, what it allows and how it would run on your PC before anything is downloaded. Once you accept the maker's terms, the file is downloaded and checked; the model sits under "Added by you", marked as not checked by GlirOS.
Research: while it works, the card shows a light circling the sources found and a moving progress bar, so you can see it hasn't stopped; while it waits for you to finish chatting, the picture stands still. When the remote connection from a phone drops, it now says so in words.
Research reports have pictures: each section opens with a moving header with its title and key words, and a section that explains how something works gets its steps, numbered and in order, with a light passing from step to step. All made on your PC; in print they stand still.
Tools
Charts 1.2.3 and Map 0.3.2: numbers in a research report are written the way your language writes them (9,600 or 9 600, never "10k"), categories are shown as horizontal bars, and map pins that sit on top of each other are moved apart; on a phone, chart words stay readable and long names sit above their bars.
Charts and maps in research reports, and better prompts
Desktop app
Research reports show their numbers as a chart and the places they name on a small map, with quiet moving backgrounds behind the headings (still when printed). Every number in a chart is checked against the source it comes from, and the map says its positions are approximate. Nothing extra goes online: the map is drawn on your PC.
Save as PDF on a research report now starts with a cover page (the question, the short answer and how it was made, on a quiet background) and a contents page where every entry is a link with a line on what is inside.
Ask "refine this prompt: …" (or "lag en god prompt av dette: …") and GlirOS helps you write a better one. A larger model gets GlirOS's own short prompt guide; with a small model, GlirOS asks a few questions in a form and writes the prompt itself. Use this prompt puts it in the message box, for you to change or send.
Web version
Ask "refine this prompt: …" (or "lag en god prompt av dette: …") and GlirOS helps you write a better one. A larger model gets GlirOS's own short prompt guide; with a small model, GlirOS asks a few questions in a form and writes the prompt itself. Use this prompt puts it in the message box, for you to change or send.
Tools
Translator 1.6.0 and SCORM 1.6.0: subtitles they make can be read aloud with Speak, which opens GlirOS's Speak the subtitles with the file; you pick the voice. Charts 1.2.1, Map 0.3.1 and Animation 1.6.1 make the report's figures and backgrounds.
A task in your lists can have a day it is due, and a time: say it ("put 'pay the invoice' on my list for Friday") or set it with the calendar button next to the item in the Tasks view. Dates show the way your language writes them, and an overdue task is marked.
"Remove the background" on a picture with a green or blue screen: GlirOS takes the screen out itself (Smart cut) and gives the picture back as a PNG with a see-through background. A picture with no even screen is told so plainly. The buttons over a picture you add now sit in one row in its corner.
The background remover is here: an AI add-on (Settings > Tools > AI add-ons, 474 MB) that cuts a person or thing out of a picture or video without a green screen, on Media Studio's timeline. Under a second a picture on a graphics card, about 5 seconds on the processor (you're asked first).
Web version
A task in your lists can have a day it is due, and a time: say it ("put 'pay the invoice' on my list for Friday") or set it with the calendar button next to the item in the Tasks view. Dates show the way your language writes them, and an overdue task is marked.
"Remove the background" on a picture with a green or blue screen: GlirOS takes the screen out itself (Smart cut) and gives the picture back as a PNG with a see-through background. A picture with no even screen is told so plainly. The buttons over a picture you add now sit in one row in its corner.
Your appointments and notes are kept on this PC in every chat, and go into the encrypted backup. Reminders for appointments come as a notice even with no chat open, and appear in Settings > Tools > Timers.
Before the AI deletes an appointment or a note, you are asked, and only your click counts. After the AI has read a web page, a document or another service in the same message, it also asks before adding to your calendar or notes.
With Memory off, "remember that …" offers once to keep it as a note in My notes instead.
When an answer says something was done but the change it needed was refused (a tool turned down a move, an add or a "mark as done", and it never worked later), GlirOS says so in one line under the answer: "Not done: … was refused".
Research writes its report in the language you asked in, plans for today's date, and in a comparison first finds out what each option is. Its findings state the facts instead of "the text mentions". When a web search fails it is tried again; when no search works the research stops and says why, and a question with no answer says whether nothing was found, the pages couldn't be read or none of them answered it. A research about GlirOS reads GlirOS's own guide too.
Web version
When an answer says something was done but the change it needed was refused (a tool turned down a move, an add or a "mark as done", and it never worked later), GlirOS says so in one line under the answer: "Not done: … was refused".
Tools
My calendar: your own appointments on this PC, with a month, week and day view, repeats, holidays, reminders and .ics import and export. Ask in your own words ("put the dentist in my calendar on Tuesday at 12"); a week you ask about is shown as a card under the answer.
My notes: short notes you keep on this PC, found again by any word or #tag, pinned first, with .txt and .md in and out.
My calendar shows your country's holidays, or all special days while Calendar & seasons is on (the Special days switch), your reminders from Timers & triggers, and trips from Trip planner: "Add to My calendar" puts a trip in, and later changes to its dates follow by themselves.
Timers & triggers reads dates written year first too (2026/04/03).
GlirOS checks what an answer says was done. When a local model says it saved, added or moved something but the step that does it never ran or was refused, GlirOS tells the model, and it does the step or says plainly it wasn't done.
When a tool refuses a request, GlirOS says why and how to fix it: a date written as 14 October 2026 becomes 2026-10-14, a place like "Bergen, Norway" is tried as "Bergen", and an id that doesn't exist is looked up first. Small models now try again instead of giving up.
A research you planned, or one that was running, comes back after GlirOS restarts: its chat shows the plan to start, or says it stopped when GlirOS closed, with the plan to start again.
When a model has already looked up the list and still uses an id that doesn't exist, GlirOS points it to that list (or to the ids the tool itself names) instead of asking it to look again.
Model test: when you continue a test that stopped before the end, the page now shows how far it has come, with a Stop button. At the end the result is saved as usual.
A research planned before GlirOS kept plans opens again: its chat offers to make the plan again from its question, and nothing starts until you press Start. Settings > About lists each change once.
While a model on this PC reads a picture, the chat says "Reading the picture …". Reading a picture can't be stopped half way, so Stop takes effect once it has been read. The model test gives pictures more time and waits until the model is free before the next item.
Web version
GlirOS checks what an answer says was done. When a local model says it saved, added or moved something but the step that does it never ran or was refused, GlirOS tells the model, and it does the step or says plainly it wasn't done.
When a tool refuses a request, GlirOS says why and how to fix it: a date written as 14 October 2026 becomes 2026-10-14, a place like "Bergen, Norway" is tried as "Bergen", and an id that doesn't exist is looked up first. Small models now try again instead of giving up.
When a model has already looked up the list and still uses an id that doesn't exist, GlirOS points it to that list (or to the ids the tool itself names) instead of asking it to look again.
When a tool has a newer version, the app says so itself: Update available on the tool in Settings > Tools, a line above the list with Update all, and one quiet notice in the bell. The first time such a tool is used in a chat, the AI mentions it in one short sentence. An update keeps the tool's saved data and a paid licence.
After a place search that found nothing or only wrong suggestions, just write the address again ("Storgata 1, Oslo" or "try Storgata 1"): GlirOS searches again by itself and shows it on the map, even with a small model.
Web version
The guide has a new chapter, Keeping tools up to date. In the browser, tools always come from the site, so they are always the newest.
After a place search that found nothing or only wrong suggestions, just write the address again ("Storgata 1, Oslo" or "try Storgata 1"): GlirOS searches again by itself and shows it on the map, even with a small model.
Replace text in a video or a GIF: in Media Studio, Replace text on a clip opens its first frame in the Picture tool; mark the text, type the new words, and Replace in the clip puts them on the whole clip, with the sound kept. It works where the camera and the text stand still; GlirOS checks that first and says so when they move.
Replacing text in pictures keeps busy backgrounds cleaner: the fill no longer copies a neighbouring line into the gap.
Web version
Replacing text in pictures keeps busy backgrounds cleaner: the fill no longer copies a neighbouring line into the gap.
Replace text in a picture: in the Picture tool (+ Add a control > Replace text), find the text or drag over it, type the new words and press Replace. The old words are taken out and the new ones drawn in the same size, colour and angle, as a new picture. You can also ask in the chat: replace "SALE" with "SALG". Pictures only, on this device.
Media Studio: a picture open in One file has Replace text, which opens it in the Picture tool.
When a tool finds nothing and you answer "yes, try again" (or give the words again), the chat now makes the call again instead of asking once more.
Map 0.2.1: a search that finds nothing keeps your words in the search box and says what to try; a near spelling is offered as "Did you mean?". Media gallery 1.7.0: full screen uses the whole width, with a size slider for the thumbnails. Update them in Settings > Tools.
Calmer text: the interface now uses a few set text sizes instead of many slightly different ones.
A research report can be saved as a PDF: Save as PDF under the report prints it in its light print layout, with the contents first.
Headings in answers are easier to follow: a title stands out, and each level below it is clearly smaller or quieter.
Web version
Replace text in a picture: in the Picture tool (+ Add a control > Replace text), find the text or drag over it, type the new words and press Replace. The old words are taken out and the new ones drawn in the same size, colour and angle, as a new picture. You can also ask in the chat: replace "SALE" with "SALG". Pictures only, on this device.
When a tool finds nothing and you answer "yes, try again" (or give the words again), the chat now makes the call again instead of asking once more.
Map 0.2.1: a search that finds nothing keeps your words in the search box and says what to try; a near spelling is offered as "Did you mean?". Media gallery 1.7.0: full screen uses the whole width, with a size slider for the thumbnails. Update them in Settings > Tools.
Calmer text: the interface now uses a few set text sizes instead of many slightly different ones.
Headings in answers are easier to follow: a title stands out, and each level below it is clearly smaller or quieter.
Research in a project: in the project's ⋯ menu, Research… takes a question, makes a plan for you to check, then works in the background: it searches the web, reads pages and the project's own documents, checks its draft against the sources and writes a report with a menu, numbered sources and appendices. Three levels: Overview, Thorough and Deep dive.
While a research runs you can go on chatting: with a local model it waits while your answer is written and then goes on; a card shows each step and a dot for each source it found. Claims that only a weak source supports go under "What we could not confirm".
The report is kept in the project (and in its linked folder), every chat in the project can find it, and it opens in the side panel or exports as a zip.
After a tool call, small models are now told to make the next call when the task needs another step, and to fix a failed call (or say plainly it couldn't be done) instead of finishing at once.
When you ask for an image editor, the chat now points you to GlirOS's own: the picture tool and Media Studio.
Model test: maths answers get more room before they are cut off, and a multi-step tool task can now take its second step, so results from before 0.1.69 aren't directly comparable for maths, tools and connectors.
The guide has two new chapters: Connectors (MCP), and Research. Ask the chat about either.
Web version
After a tool call, small models are now told to make the next call when the task needs another step, and to fix a failed call (or say plainly it couldn't be done) instead of finishing at once.
When you ask for an image editor, the chat now points you to GlirOS's own picture tool.
The guide has two new chapters: Connectors (MCP), and Research. Ask the chat about either.
Small local models often write a tool call as "weather.forecast" with a call name added; GlirOS now reads that correctly, so the call runs instead of the model saying the tool isn't available.
Model test: the run without GlirOS's help now lists the test tools the way it asks for the call, and a plan follows the steps the task asks for. Results from before 0.1.68 aren't directly comparable for tools and plans.
The bell badge on GlirOS's taskbar button comes back when you open the window again from the tray; before, only the tray icon showed it.
Model test: a plan that starts with "Write the script to …" is no longer marked as writing first, and Compare says how far a Quick test can swing an area.
Gemma 4 models' own way of writing a tool call is now understood, so their calls run.
Model test: tool calls without GlirOS's help are read the way a plain app reads them, connectors are tested as the app uses them, and several calls in one reply all run.
GlirOS's code check no longer reports SVG filter and animation tags (such as a blur or a glow) as errors, so a drawing that was right isn't sent back to the model to be "fixed".
Small models that misspell GlirOS ("GliOS") are told the right spelling, and the name is put right in their answers, on screen and read aloud.
Web version
Small local models often write a tool call as "weather.forecast" with a call name added; GlirOS now reads that correctly, so the call runs instead of the model saying the tool isn't available.
Gemma 4 models' own way of writing a tool call is now understood, so their calls run.
GlirOS's code check no longer reports SVG filter and animation tags (such as a blur or a glow) as errors, so a drawing that was right isn't sent back to the model to be "fixed".
Small models that misspell GlirOS ("GliOS") are told the right spelling, and the name is put right in their answers, on screen and read aloud.
A clip on the timeline now has the full colour grading in its own Colour and look part: Warmth, Tint, Fade, Teal & orange and the ten ready looks, the same as for a single file.
Web version
A clip on the timeline now has the full colour grading in its own Colour and look part: Warmth, Tint, Fade, Teal & orange and the ten ready looks, the same as for a single file.
A repeating timer set in your own time zone now stays at its time there, even in the weeks when this device changes to or from summer time on a different date. The reminder says which place's time it is, for example "08:00 (Tokyo)".
Four new models on the list for big PCs with a 16 GB graphics card or more: Qwen 3.8 27B, Qwen 3.6 27B, Qwen 3.6 35B-A3B and Qwen3 Coder 30B. They are offered in Settings > Models but not suggested until GlirOS's benchmark has tested them.
A plan's progress bar shows properly again when the plan has five steps or more.
Web version
A repeating timer set in your own time zone now stays at its time there, even in the weeks when this device changes to or from summer time on a different date. The reminder says which place's time it is, for example "08:00 (Tokyo)".
A plan's progress bar shows properly again when the plan has five steps or more.
The guide is easier to read: its longer chapters have paragraphs and headings, and when you ask GlirOS about something, it now picks the part of a long chapter that your question is about.
Web version
The guide is easier to read: its longer chapters have paragraphs and headings, and when you ask GlirOS about something, it now picks the part of a long chapter that your question is about.
Settings > General has Your time zone: automatic (this PC's) or one you choose. GlirOS gives the time and the date in it.
Say where you are ("I'm in Tokyo") and that chat follows you: the time and the date are given there, with the zone named, and GlirOS offers once to save it as your time zone.
Web version
Settings > General has Your time zone: automatic (this browser's) or one you choose. GlirOS gives the time and the date in it.
Say where you are ("I'm in Tokyo") and that chat follows you: the time and the date are given there, with the zone named, and GlirOS offers once to save it as your time zone.
Tools
Date & time 1.1.1 tells the time in any place ("what time is it in New York?") and turns a clock time from one place into another, on your device without going online. Calendar & seasons and Timers & triggers now use your time zone too.
GlirOS now checks its signed model list on tools.gliros.com once a day, so new models, fixed settings and test results reach Get more models without a new version. The models still download from Hugging Face, checked byte for byte; nothing about your PC is sent, and a list that isn't GlirOS's own is never used.
Get more models says when a new model on the list fits your PC (once, never a pop-up), whether GlirOS's benchmark has tested a model, when a model needs a newer GlirOS, and when one was taken off the list and what to try instead. Check now looks at the list at once.
Each model answers with its maker's recommended settings
Desktop app
Settings > Benchmark can now run a model with its maker's recommended settings or with GlirOS's earlier settings, besides the fixed setting, so you can see which answers better on your PC. A run with varied answers repeats each question three times.
The models in Get more models now answer with the settings their makers recommend (how varied the words are, and a guard against repeating itself), with different settings for thinking and for a quick answer. Settings > Models says where each model's settings come from, with a link to the maker's model card.
Loading a model sets the think button the way its maker recommends: Qwen 3.5 2B starts with thinking off (it can get stuck thinking), Qwen 3.5 4B and 9B with it on. You can switch it until the next time a model loads. With gpt-oss, thinking off now means a short line of reasoning (it can't stop reasoning completely).
Tools
Translator 1.5.1: Save in the chat is back (each finished language is kept in Files and versions), and it says when a language was left out because it is the text's own language.
More room for picture models, and SVG drawings as pictures
Desktop app
An SVG drawing in an answer can now be opened as a picture in the picture tool (Edit as a picture) to crop, adjust and save as PNG. The picture tool's Open a picture takes .svg files too.
A few small text weights were evened out to match the rest of the design.
Web version
The browser models that see pictures have a small memory. GlirOS now gives them shorter instructions, so a chat with a video and a tool keeps more of the conversation and the model's actions run reliably.
Speak the subtitles: a voice on your PC reads a subtitle file (.srt or .vtt) line by line at the right times, and the result is one sound track for your video. Lines that are too long are spoken a little faster to fit. Nothing leaves your PC.
You find it in Sound Studio (Speak subtitles, with one file per speaker if you like, each on its own line), in Media Studio's text panel once it has subtitles, and in the chat: ask GlirOS to read a subtitle file aloud.
GlirOS guesses the language from the words and asks you when it isn't clear, and it only offers voices that speak that language.
Fixed: in Sound Studio, choosing a sound under From the chat now adds it.
Web version
In the browser version the subtitles are read in English with the natural voice, at its own speed; the desktop app can also speed lines up to fit.
On a phone, the line beside the send button no longer squeezes into a tall column: it gets its own line under the buttons, and the small-helper reminder is left out there, since the line under the box already says answers can be wrong.
A model that thinks before it answers could use all its room thinking, so the answer never came. GlirOS now stops the thinking in time and has the model answer from what it has.
If an answer still ends while the model is thinking, Continue now writes the answer instead of showing an error, and the "this chat is too long" message only appears when that is really the problem.
Models that can see pictures now load their picture part only when the first picture comes. Chats without pictures get that memory for a longer conversation, and GlirOS keeps at least room for about 8,000 tokens before it moves work off the graphics card.
Transcripts: ask "transcribe it" about a video or recording from earlier in the chat, see the whole text under the answer, and download it as .srt, .vtt or plain text. The model knows GlirOS can transcribe and can do it itself, and a transcript made once is used again.
Media Studio's text panel can use the chat's transcript, transcribe the clip, take subtitles from your PC and download them; Export can save the transcript as .srt, .vtt or .txt.
Web version
On a phone: the Back button closes Settings and panels first, the screen stays on while an answer is written, and a remote session says clearly when it can't reach your PC. A line under the message box reminds that answers are made by AI, with a link to learn more.
A long message, a pasted text or a file can take a while for a model on your PC to read before it starts answering. The chat now says how far it is: "Reading your message … 57 %".
Settings > Models shows how the loaded model uses memory: the model itself, the conversation and the work it needs. Under Advanced there is a test setting for 8-bit conversation memory, which uses about half as much for long chats.
Face swap in Media Studio takes up to five pictures of the same person (More pictures of this person). Their faces are combined, so the result looks more like the person and less like any one photo.
Swapping only one person in a video keeps to them better: GlirOS learns their face over the clip and holds on to it through a blink or a turned head, without taking anyone else. A cut to a new scene starts the check afresh.
Sound Studio, a mixer beside the chat: put voice, music and effects on lines, set each line's volume and pan, let the music duck under the voice, add markers, and record onto a line while the others play. Export mixes it into one file in the chat (WAV, MP3, AAC, FLAC or Opus), or straight onto Media Studio's sound line. Open it from a sound's card, with /sound, or by asking.
Clean up a recording in Sound Studio: take out rumble and steady background noise, soften sharp s sounds, shorten long silences, and bring it to an even loudness for speech, a podcast, music or broadcast (measured first, then set exactly). The clean copy comes into the chat and takes the block's place; the original stays.
Web version
Sound Studio works in the web version too: the browser mixes the lines (with pan and ducking) and evens out the loudness, into a WAV file. Noise removal, the other formats and exact loudness need the desktop app, and the panel says so.
An error message now has a Report button. It opens Send a report with the error quoted (personal paths and keys taken out) and room for what you were doing; nothing is sent until you press Preview and Send.
Web version
The web version's page now refuses any script that isn't GlirOS's own, so even a mistake that let foreign code into the page couldn't run it. Tools and previews keep working in their own frames.
A document tab kept with a chat now opens again after a restart, instead of saying the document is no longer in the chat.
Web version
In the web version, tools and code previews now run in a frame of their own that can't reach your chats, files or keys. That lets the page itself refuse any script that isn't GlirOS's own; this version checks that rule and notes anything it would block in Send a report's diagnostics, and the next one turns it on.
Media Studio now belongs to its chat: the video, the parts you set and the timeline come back when you open the chat again, also after a restart. Chats with a Media Studio setup have a small film mark in the chat list, so a project's chats work as your list of media projects. A job finishes in the chat it was started in, even if you move on, and an incognito chat keeps nothing.
Remote access now lets in only browsers you have allowed. The first time you open your PC from a browser, GlirOS on the PC asks whether to allow it (with a fingerprint the browser shows too), or you make a code there and type it in the browser.
Remote access gets ready to let in only browsers you have allowed: Settings > Remote access > Allowed browsers. Each new browser is allowed once at the PC, when GlirOS asks or with a code you make there, and you can remove a browser at any time. Your phone keeps working as before until the Open page is updated to ask.
SVG drawings, faster speech recognition and safer remote access
Desktop app
Ask for SVG code ("create an SVG of…") and the model writes the drawing itself, shown as a picture with Save SVG, even when it leaves out the code block. Asking for an avatar still goes to Avatars.
Faster speech recognition on this PC: Settings > Voice can now download NB-Whisper small (Norwegian) and Whisper small (many languages) for you, checked byte for byte. It is optional: the built-in speech recognition still works without them.
Remote access is safer: your PC now lets in only connections your GlirOS account has signed, so even someone in control of the relay can't open a session on it.
A security review of the desktop app and the web version. Through remote access, connectors can no longer be added or changed (a local connector starts a program on the PC), so that is done on the PC itself. Tool pages always run in a sandbox of their own, previews respect Settings > Security's network block, a restored backup can't bring files a backup never holds and brings program connectors back switched off, and a cloud server moved to another address no longer keeps its key.
Web version
Files are checked before they are opened: a Word, PowerPoint or e-book file that would unpack to far more than its size (a "zip bomb") is refused instead of freezing the page, and source links in answers are only ever web addresses.
Colours that are easier to read: the delete button in the dark theme has dark text on its red (white was hard to read), and in the light theme the green, amber and red used for small status words, and the grey of code comments, are a little darker. A few colours from outside GlirOS's palette are gone, and the waiting tasks have a calm outline instead of a hard bar.
The guide's Media Studio chapter explains face swap's Fine-tune (How it is pasted, Restorer amount, Keep anything in front of the face), the result's Save as… and Keep working on this, saving a GIF as it is, and why Media Studio empties when you move to another chat.
Face swap in Media Studio looks much more like the source face: it now works the way the old face swap web tool did, and the face restorer is mixed in at half strength. On our test pictures the likeness went up on every face. The smoother way is still there as a choice in Fine-tune.
Face swap settings are used even with Fine-tune closed (before, closing it lost them), "Keep anything in front of the face" is off unless you turn it on and then has a level, and the restorer has an amount. When the chosen person isn't recognised (closed eyes, a turned head), Media Studio says so.
A result shows in Media Studio as soon as it is made, with Original and Result to compare and Save as… to keep the file. Keep working on this keeps your settings, with Back to the original. Export can save a GIF as it is, and a GIF put on the timeline becomes a video copy so it plays there. Moving to another chat empties the panel, so nothing from an incognito chat stays.
Asking for a game, even "with cool animations", gets a page you can play in the preview, not an Animation video. Code files attached in the chat are taken even when they have no text in them (a canvas game's index.html), go into Files and versions so the Code workspace shows them, and the model is told the chat's code files instead of asking you to upload them. Make video on a page split into several files hands Animation one file with everything in it.
Switching a setting on (such as a tool) no longer stops at "Could not save the settings file: Access is denied" when a virus scanner or a sync program is holding the file for a moment. If it still can't be saved, the message says where the file is and what may be holding it.
Moving pictures in Media Studio, and Animation's Code template
Desktop app
Media Studio shows GIFs again, before and after a job (they stayed blank). A face swap or face age on a GIF or an animated WebP now changes every frame and keeps the GIF or WebP, at its own size and speed, instead of making one still picture.
Videos recorded in a browser (WebM with no length or frame rate in the file) keep their length through the AI add-ons, and Media Studio shows how long they are.
Web version
In the browser version, tools that work with the chat's files (Animation, Avatars, the Website builder) can read and write them again.
The GlirOS Helper (GPU) asks the browser for the fastest graphics card, and Settings > Models names the one it got. On graphics built into the processor (shared memory) it warns, isn't loaded by itself at start, asks before loading, and the first start recommends Lite. On a laptop with two graphics cards, Windows Settings > System > Display > Graphics can set the browser to High performance.
Tools
Animation 1.4.1 has the Code template: any animation as one HTML page with a canvas, recorded frame by frame into a video, with Describe a change. Make video works on code cards in the chat and in the Code workspace's preview. Map 0.2.0 takes places and asks you to pick one; Timetables, Picture to 3D, Handcraft, Calendar & seasons and Website builder are updated.
Animations no template does: ask for one ("an animation that counts from 1 to 10") and the model writes it as code, one HTML page with a canvas, like a small game, which you can play in the preview. When your Animation tool has the new Code template, Make video under the code records it into a video.
With the globe on, a short reply to the model's own question ("10", "green") is no longer searched on the web. Real questions still are.
Settings > Benchmark has the presets pack: it tests one model on GlirOS's own suggestions, every tool's first examples, and whether it calls only the tools that are there, with GlirOS's help and without. Its own table and an export, outside the benchmark's scores.
Tools
Avatars 1.1.0 shows drawings a strong chat model writes itself as SVG ("draw a sleepy wizard as SVG"), made safe first, with Save SVG and Save PNG. An SVG written straight into an answer now shows as a picture too, with Open in Avatars. Made-up characters only, never a likeness of a real person.
Memory across chats, off until you turn it on in Settings > Memory. Say "remember that I'm vegetarian", or keep a fact GlirOS offers under an answer. "What do you remember about me?" lists them, and Settings shows everything to edit or delete, with a paste from ChatGPT. Never used in an incognito chat or by the Helper, and a cloud model is named the first time it gets them.
Recipes: when a plan is finished, Save as recipe keeps its steps. Run it again with new input from Settings > Recipes, or write "run my recipe" and its name, and the model plans from the saved steps. Share a recipe as a file, and import one: its steps are shown before it is kept.
How this answer was made now shows how soon the first word came and the speed. A chat near the model's limit says so under the answer, and Settings > Models says how much a loaded model can read. Sums read numbers the way your language writes them (1,234 or 1.234 or 1 234), and "15 % of 2,400" is worked out by the calculator.
Tools
Every GlirOS tool now has its name, description and example requests in French, Italian and German too, so a small model finds the right tool in those languages. Tools can now hand a place or other small piece of data to another tool, and ask you to pick a place on the Map.
New in the tool catalogue: Avatars. Cartoon avatars in three free art styles, from a few choices or your own words ("short red hair and glasses, smiling"), saved as SVG or PNG for a profile picture or a website. They are generic cartoons, never a likeness of a real person.
A Helper that knows its job, and faster small models
Desktop app
Small models answer sooner in long chats: the start of what they read stays the same from turn to turn, so it isn't read again each time. With many tools on, a small model now gets the calls of the right tool far more often (583 of 594 example requests, from 209), because GlirOS also knows each tool by its examples and names in every language.
Fixed: with a small model, a big job whose plan couldn't be read was refused with a technical message and nothing was sent (seen on the Norwegian Lofoten suggestion). Messages like that from inside GlirOS now say in plain words what to do. Tables in answers keep their words whole on a phone, the model menu fits narrow phones, and a TV or 4K screen shows GlirOS larger.
Web version
The GlirOS Helper now sticks to GlirOS. Its suggestions are GlirOS questions, and for anything else GlirOS answers first: it offers a bigger model here, Choose a model, or Ask the Helper anyway. A job one of your tools does gets Open (or Switch on) for that tool, because the Helper can't use tools itself.
Tools
Clothing design 1.13.1 warns when a design saved in a brand-new chat wasn't kept, and says to write a message first. Tools now hear whether a save was kept, so they can say so at once.
New in the tool catalogue: Map (Norway). Kartverket's map in the side panel, search for places and addresses, pins the model sets (its own coordinates are always shown as approximate), your own marks with the nearest address, and distances in a straight line. Say "show on the map".
A plan now stays one plan: when the model makes the same plan again it goes on with the one it has, and only one plan runs at a time. "Finish the tasks" goes on with the open plan, and news for the job ("my brother is coming too") is added to it instead of starting over. Small models get what the job has worked out so far with each follow-up.
When you ask for something new while a plan is still open, GlirOS asks first: end it and start this, add this to the plan, finish the plan first, or keep it paused. Only a plan that is running shows ▶ and moves the bar, and Continue is only offered when the plan can go on.
When the model calls a tool that isn't switched on, or says it opened one, GlirOS says plainly under the answer that nothing ran, with a button to switch it on. Tool calls no longer show as code, "[END]" marks and copied dates are taken out of answers, and typing "no" to the project suggestion just closes it.
Under every answer, "Report this answer" opens Send a report with the answer quoted, so you can tell GlirOS support about a wrong or harmful answer. You see exactly what is sent, and nothing goes until you press Send.
"No need, just open the tool for clothes" now opens Clothing design: a "no" only counts for the part of the message it belongs to, and "the tool for …" finds the tool that matches. When an answer names a tool by its code name, it shows the tool's own name as a button that opens it.
Code catalog 1.3.1: when a page says others see it (a live chat, other players, votes so far) but everything stays in your own browser, GlirOS says so under the answer: it needs a server or a real-time service in between, and until then the page should be labelled as a demo.
When you ask for a database (a customer or member register, stock) or for something several people use at once (a live chat, a multiplayer game, editing together), GlirOS now has the model say plainly what a page alone can't do, and label any demo that only works in your own browser.
Tools
Code catalog 1.3: ready-made PHP files for a log-in, members' pages, a forum and comments, each with the files it needs, and a "What you need to do" file with the steps and a test list for your own web host. GlirOS writes them and checks them, but can't run them here: test them on your host before you rely on them.
For trying GlirOS's remote access on a new server before it goes live: the desktop app can be started with a test relay, and Settings > Remote access says so plainly at the top. Starting it normally uses the usual relay.
Speech recognition on this PC: a very short sound no longer risks a crash, and stopping it while it finds out which language you spoke takes a moment instead of seconds.
Tools
SCORM 1.3.2: AICC courses (for example Questionmark's exports) can be merged into one AICC course with a choice page. The chosen course reports to the LMS by itself, exactly as it would on its own, and nothing about the learner passes through the page. AICC and SCORM courses can't be mixed in one merge.
Under an answer with code, GlirOS now says when the code is written but not run here (PHP, Python, SQL), and when it isn't safe as written: a password checked in the page, a key in page code, a home-made card form, or a form that says "sent" with nothing behind it.
Tools
Website builder 1.5: a Publish step uploads your site to your own web host (desktop app), shows what will change first, and asks before anything is sent. Pages that would need a server (log-in, forum, shop payments, bookings, live data) are offered as ordinary pages instead of fakes.
Speech recognition on this PC: with a Whisper model in the models\stt folder, the microphone button and conversation mode write down what you say on all your processor's cores instead of one, several times faster. Settings > Voice shows the models and has Try both, which times this PC against the page on one sentence. Nothing leaves the PC, and the page's own Whisper still takes over if needed.
Web version
When you ask for something a page can't do by itself (log-in, a shop, sending e-mail, live data, a chatbot), the model is told to say so and build the safe version instead of faking it. Plan steps whose code GlirOS can't run here (PHP, Python, SQL) are marked "written, not run".
Settings > Remote access has the switch "Let my web app use this PC's model". Videos: Small, Normal and High now give about the same quality on every video encoder (measured on real PCs), and Settings > Media says when an older Nvidia driver keeps the graphics card from making videos.
Web version
My PC: in GlirOS on the web, Settings > Cloud models > My PC finds the PCs linked to your account and lets you use a PC's model from any browser. Only model requests go there, end-to-end encrypted through GlirOS's relay; the PC's chats, files and tools can't be reached. The PC must be on, with "Let my web app use this PC's model" switched on (GlirOS Plus).
Tools
SCORM tool: it recognises Articulate Rise Connected Packages (the course is loaded from Articulate's servers), offers only their LMS title for translation, and merges them with other courses; learners then need access to Articulate's servers.
Website hosts (Settings > Tools): add your own web host over SFTP or FTPS, test the connection, and the Website builder can upload your site there. GlirOS asks before each upload, sends only what changed, and never removes files it didn't put there. The password is kept in Windows Credential Manager. Plain FTP isn't offered, because it sends the password unencrypted.
Media Studio's timeline: drag a clip down onto a layer line, crossfades show in the preview, the colour key stays in the preview at every frame, and Colour and look, Crop and the colour key have Undo and Reset. A clip opened in One file goes back in its place with your changes, and One file's Adjust can be opened from a clip.
Tools
Sports lookup: questions about Toppserien, OBOS-ligaen, Allsvenskan and Superligaen go to the lookup, and the past winners now cover all 15 competitions, Serie A back to 1898.
Export: each sharing preset says what its limit is for (Discord's free 20 MB, the e-mail sizes, WhatsApp as a video or a document), with the limits checked again on 8 October.
Media Studio's timeline fills the screen the way One file does: the preview with the chosen block's settings beside it, and the rows under them. A line's volume now floats over the rows, and a new To One file button makes the timeline and opens the result in One file, ready to export.
Tools
Website builder: 38 kinds of sections (tabs, maps, opening hours, forms, a countdown, before/after and more), two new looks, forms that send by e-mail or with a small contact form on your own web host, importing a site you already have, a full-screen editor with drag and drop, and photos made in several sizes so pages load faster.
Settings > Tools no longer shows English notes on a Norwegian or Spanish page: a tool's note shows in your language when the tool has it.
Export's Vertical preset can keep the whole picture: Blurred sides puts it in the middle of the tall frame over a blurred copy of itself. Or crop the middle, or follow the person. "Show where the apps put their buttons" marks the part of the frame TikTok, Reels and Shorts leave free.
WhatsApp can also be sent as a document in full quality. Export can even out the volume, and the sharing presets leave out the location and camera data the file carried.
Tools
Clothing design: five more garments: a pencil skirt, summer trousers, a tunic, a zip hoodie and a deerstalker hat.
Clothing design: children's sizes by height (62-152 cm) with child mannequins in 3D, a baby romper and a bib.
Website builder: a blog, products, events, team and FAQ lists; the site in more than one language; search; a section with your own code; and nothing you changed by hand is overwritten without asking.
Sports lookup: "tell me about Molde" is answered on your device for the clubs of the covered leagues and Norway's national teams, in English, Norwegian and Spanish.
Media Studio's Export part has presets for the usual places: Discord, WhatsApp and e-mail (made to fit their size limit, checked after), YouTube, vertical 9:16 for TikTok, Reels and Shorts, a presentation, sound only, GIF, a picture, high quality, small, lossless, or the original without re-encoding. A line says about how big the file will be and how long it takes; Details let you change the format, resolution, codec, quality and frame rate.
Or just say it in the chat: "make it fit Discord", "for YouTube", "make it vertical for TikTok", "send it by e-mail".
Tools
Clothing design has a tank top, leggings, swim shorts and a circle skirt, each with its own choices, pattern pieces, sewing steps and 3D preview. The pockets panel now shows the places of the garment you are making.
Write what you want done with your green-screen videos ("remove the green screen, put beach.jpg behind, name bars, fade to black, separate files") and GlirOS shows the plan for each video, asks once for anything missing, and makes them all. Each finished video comes into the chat and can be opened in the timeline.
Tools
Handcraft hands sewing a hoodie, trousers, shorts or a flat cap to Clothing design.
You can write while GlirOS is working: Enter puts the message in a small queue above the message box, and it is sent by itself when GlirOS is done. Each waiting message has Send now, Edit and Remove, and the queue is kept if you close GlirOS. "Stop" stops, and "how far are we?" is answered at once.
Tools
Clothing design can make a hoodie (raglan sleeves, a hood, a kangaroo pocket), pull-on trousers or shorts, and a flat cap. 3D shapes, Picture to 3D, Charts and Teleprompter can save a short video: a turning model, a chart growing in, or the script scrolling.
Compare views in Media Studio, and GlirOS Plus status
Desktop app
Media Studio's View menu shows before and after with a divider, marks clipped blacks and whites, or shows the picture in black and white. Waveform, parade and vectorscope sit beside the colour graph, and "make it look like beach.mp4" matches the colours to another file. Views change only the preview.
Settings > Remote access shows your GlirOS Plus: a free trial with the days left, active, or how to start it. Everything is started and managed at home.gliros.com; nothing is bought in the app, and everything else in GlirOS stays free.
Opened from a browser through your account, GlirOS shows a green pill with your PC's name at the top: press it for the connection and Leave. On the PC, a pill says when it is in use from a browser, with End session.
Tools
Clothing design can make a thin sweater: long sleeves, ribbed cuffs and waistband or plain hems, and shorten the sleeves as you like. The T-shirt's longest sleeve is now called three-quarter.
Handcraft now does sewing: a tote bag, a cushion cover and an elastic-waist skirt, with calculators for cutting, fabric and elastic. A garment from Clothing design can be followed step by step in Handcraft.
Sharing goes through your GlirOS account only: the old share key field is gone from Settings. The benchmark now also measures how reliably a model calls tools, recovers from an error and avoids repeats.
Tools
Clothing design shows your T-shirt in 3D on a mannequin: turn it, pick a pose, see through the body, set your own lengths, and save it as a 3D file. The sewing booklet's cover shows it too.
Chess can share a game: GlirOS asks first, then gives you a watch link and an invite. Your friend plays in their own GlirOS or on the web page, spectators follow, and the clocks are kept by the server.
A tool that needs your own key for a service asks for it in GlirOS's own window. GlirOS keeps it in Windows Credential Manager (or this browser) and uses it only for that service; the tool and the AI never see it. Settings > Tools lists the keys with Forget.
When a tool fills the screen, the chat floats over its left side: the last messages as boxes, a field to write in, the answer arriving in the newest box. Click a box to bring the chat back.
Tools
New in the catalogue: Chess, to play and learn against the engine, the chat model or a friend here, with a coach, clocks, 3D and Practise my mistakes; and Clothing design, a T-shirt from a real sewing pattern to your size, with colours, patterns, prints, pockets and the sewing steps.
GlirOS's own tools can save files such as a PDF in the chat's Files and versions, and ask the chat model a short question of their own (a cloud model asks you first).
Media Studio's timeline: every clip and layer can have its own colour key. Smart cut finds the green or blue screen from the picture; Pick, Matte, an edge colour and your own levels are there when you want them. In the chat, "remove the green screen" does it on the chosen layer.
Each clip and layer keeps its own sound, linked so it moves with it, with mute, volume and fades; each line has its own volume. Detach sound puts it on the sound row on purpose, Sync again puts it back in step, and every block shows its waveform.
Media Studio's Adjust part has new Warmth, Tint, Fade and Teal & orange sliders, ten ready looks (Golden hour, Vintage film, Moody, Noir and more) and a colour graph that shows red, green and blue for the frame shown, with a faint outline of the original.
Ask in the chat ("make it warmer", "give it a golden hour look") and the sliders glide into place in Media Studio, marked "Set by GlirOS" with Undo. Nothing is saved until you press Make it. For a look in your own words, the chat model chooses the values: one that can see pictures looks at the frame (with a cloud model, GlirOS asks first), and the others get the frame's measured colours.
The Translator's "Translate with the chat model" now hands its lines to GlirOS: they are translated part by part with every line checked, the panel fills as it goes, and Stop keeps what came back. With a cloud model, GlirOS asks first.
Every GlirOS tool now has a short plain line about what it does, shown in Settings > Tools and with the tool's details on community.gliros.com/tools, in English, Norwegian and Spanish.
Settings is calmer: the sections sit under a few headings, the choices used less often are under Advanced at the bottom of each section, and the tool list has a filter and scrolls by itself. The calendar and your country are in General now.
If GlirOS was closed while it was working, nothing goes on by itself at the next start: a card says what was stopped and offers what fits, such as opening the chat or resuming a download.
Web version
Settings > About > Send a report tells GlirOS support about a problem or an idea. You see exactly what is sent, and nothing goes until you press Send.
Tools
The tools got a layout pass: the same header, cards and help text everywhere, nothing runs off a narrow panel, and clearer buttons while something is still missing.
Quitting GlirOS while an answer is being written or a file translated now asks first: keep it running in the tray and quit when it's done, quit now, or go back. The tray icon's tip also shows what the chat is doing.
Tools
Translator 1.4.0: a PDF's language is found from its text, so English is kept as a target; the browser's quick translation is checked line by line, as the model's is.
A clip on the Media Studio timeline can now have a crop that follows the main person: pick a shape under Crop and tick Follow the person. The faces are found on this PC, and the box moves with the person in the preview.
A quicker start, and the cost said before a long cloud job
Desktop app
Before a whole file goes to a cloud model, GlirOS says about how many tokens it takes, and whether that's over your limit for one answer.
After summarising a whole PDF, its overview opens on the Overview tab, where the summary is, even if you were looking at the pages.
Web version
GlirOS starts with about a quarter less to load: Media Studio, the benchmark, PDF translation and the interface languages you don't use load only when needed.
See a PDF's pages, take its text and pictures, and compare with the original in Media Studio
Desktop app
A PDF opens with two tabs: Overview, and Pages with the real pages, zoom and search. Under a PDF you add: Open, Summarise, Extract text and Translate (with how long it takes). "Copy the pictures" puts its pictures in the chat.
Media Studio: an eye on each part hides it in the preview, and holding Original shows the picture untouched. On the timeline, a line's eye switches that line off, also in the finished video.
File translation checks every answer: a translation that loses a placeholder, tag or link is asked for again, and XLIFF files (such as SCORM exports) keep their inline tags.
Ask to translate a PDF and you get a PDF back: the same pages, pictures and colours, with the text replaced where it stood. Text in pictures and on scanned pages is read on your device and translated too. Download it, or get it as text instead.
Tools
The Translator tool 1.3.0 can translate PDFs: see the pages before and after, go through what was left as it was, type your own wording and translate again.
A crop that follows the person, and a look before a plan sends anything
Desktop app
Crop a video to 9:16 or square and keep the main person in the picture as they move: Follow the person in Media Studio's Crop, or say "crop to 9:16 and follow the person". The faces are found on your PC with a small face finder (0.2 MB).
When a plan's steps will send things through a connector, such as creating an issue or sending a message, GlirOS shows what and where before the plan starts. Start it, or change the plan first.
Fixed: on some PCs, reading scanned PDFs, the voice engine and What's new didn't load in the installed app.
An app on your phone, and a fair first start on any PC
Desktop app
On a PC without a graphics card, the welcome says the model runs on the processor and why it recommends a small one. After the first answer GlirOS says how long it took to start and how fast it wrote. While a model downloads you can chat with a cloud model.
Rate answers with a thumb up or down. Settings > Models > Your ratings shows how you rated each model and which won your comparisons, kept on this device. The web version has it too.
Web version
The web version can be installed as an app on a phone, a tablet or a PC, with its own window and icon. Settings > About says how in your browser.
Tips: after a quiet while in a chat, a short tip under the last answer about a tool the chat used. Show examples gives a few things to try; a button fills the message box without sending. At most one a day; Settings > Notifications turns tips off. The web version has them too.
Scanned PDFs read more reliably: if text recognition can't start, GlirOS tries again at once, the next PDF works without a restart, and a PDF's pages with text are kept even when its scanned pages can't be read.
Paid tools say which payment company sells them (such as Paddle), and Buy goes away when a maker stops selling through GlirOS; a licence you already have keeps working until it ends.
If GlirOS was closed in the middle of an answer or a plan, a reminder next time takes you back to that chat, where you can ask it to go on. The web version does the same.
A base model (one that isn't made for chat) is told when you add or load it, with advice to look for its -instruct or -chat version.
After a benchmark run, GlirOS gives that model as much help with plans and steps as the test showed it needs, instead of guessing from its size.
Answers read as meant: when a model writes a result before the tool has answered, that guess is folded away above the real answer, and a copied tool result never shows. The web version does the same.
Web version
The web version has Settings > Backup: one encrypted file with your chats, files and versions, pictures, code workspaces, settings and, if you like, your media files. Restore it in any browser with its password. API keys are never in it.
Code in an answer is checked without running it. If GlirOS finds errors, the model fixes them in the same answer, and you see "Code checked · errors fixed". Errors in the preview can be sent with Fix with AI.
Safer long jobs: a tool call repeated again and again is stopped, and an answer is asked to finish after its time limit or, with a cloud model, its token limit (Settings > Models).
"How this answer was made" under each answer: the steps, the model, time and tokens, and what left this device. A chat's menu shows everything it sent out.
A plan step marked done is checked against what really happened, and a tool call with wrong arguments is told what the call takes.
Models on this PC write tool calls that always work: once a model starts a call, it can only name tools that are switched on and write valid JSON (Settings > Models).
Tools
Media gallery 1.5.0: mark pictures and videos, then Watermark… makes copies with your logo beside the originals (videos in the desktop app). A logo's see-through parts are kept.
Settings > Backup makes one encrypted file with your chats, files and versions, pictures, code workspaces, tools, settings and, if you like, your media files. Save it on another drive, a USB stick or a synced folder.
Restore checks the whole backup first, keeps today's data in a folder of its own until you delete it, and finishes when GlirOS restarts. API keys, licence keys and the remote access link are never in a backup.
A reminder when a backup is due: monthly by default, or weekly, every two weeks, every three months or never.
Import from ChatGPT: in Settings > Chat history, choose the export zip from ChatGPT and your conversations become GlirOS chats, with their titles, times and messages. Read on your device only; the same export twice adds nothing.
Connectors are safer: once a message has read a web page, a document or a connector's result, any connector call that changes something is asked again, with its full content and no "Always allow".
Tools
Media gallery 1.4.2: corrected examples. Tools can now save a copy beside its original in a subfolder, and put a watermark on a video in a gallery folder (desktop app).
Videos are made on your device: you see the progress with Cancel, and the video lands in the chat with Open in Media Studio. It also works when the AI uses the tool with its panel hidden, or in a scheduled task.
Watermarks: ask "put my logo on the video as a watermark" and your picture goes in the bottom right corner, 15 % wide and 70 % visible, on videos and pictures. Ask for another place or size, or drag it in Media Studio's new Watermark button. Your original is never changed.
Media Studio: Text > Title card and Name bar are made by Animation and laid over your clip.
Tall pictures in the chat are no longer squeezed: they're cropped to fit, and open whole when you click them.
Ask "how do I get an API key for Gemini?" (or OpenAI, Claude, OpenRouter, Groq, Mistral) and GlirOS answers at once with direct links to the provider's key page and its own guide.
Tools
The new Animation tool makes short videos in the chat: title cards, name bars, moving text, countdowns, growing charts, quotes, slideshows, logos and intros with lightning, with effect sounds and your music. Up to 60 seconds, 1920×1080 or upright 1080×1920 for phones.
Add a picture, such as your logo, and ask "spin it, then slide it out to the right": the AI uses it by its name. Open in Animation on a picture in the chat starts a logo animation with it.
For tool makers: GlirOS.video turns frames a tool draws into an .mp4 (H.264, with AAC sound on the desktop), and tools can use GlirOS's logo. Home-made tools ask you the first time they make a video.
Every tool now has examples: ask "what can the translator do?" for a few simple ones, then "more examples" for harder ones. Settings > Tools shows them as Try: chips that fill in the message box.
The model benchmark: the same test for every model
Desktop app
Settings > Benchmark runs the same test on any model, local or cloud, and gives a Quality score, a Speed score on this PC and one line per category, from maths and coding to tools, long chats and translation.
Coding, tools and plans are scored with GlirOS's help and without it, so you see what GlirOS adds for small models. Code answers are run and tested in a sandbox.
The full run has 161 questions in all 17 categories, including reading pictures, dictation and reading aloud, which use only what is already on this PC.
The test texts hold nothing of yours and nothing goes into your chats; chatting waits while it runs. A cloud model asks first and shows the estimated tokens, and the result shows what the provider counted.
Results can be compared side by side, exported as a picture, a file, text or a table, and imported from another PC. Compare models in the chat is now a beta.
Tools
Trip planner 1.2.1: a Search flights button on flights still to book opens Google Flights with the route and dates, after GlirOS asks.
Subtitles as blocks on the timeline, and turned texts
Desktop app
Media Studio's Text button on the timeline opens a menu: a text at the red line, or subtitles from the chat, from an SRT or VTT file on this PC, or from the clip's speech (written down on this device).
Each subtitle becomes a block of its own at its time, over the three text lines when they overlap. Move, stretch, edit or delete each one. A text holding pasted subtitles can be split into blocks too.
Texts can be turned: use the Angle chips and slider, or drag the round handle above the text in the picture (hold Shift for 15° steps). The result is turned the same way.
The user guide has new chapters on notifications, community and paid tools, and the timeline's subtitles and turned texts. The account chapter now says what a GlirOS account is for.
Web version
The web version has the same subtitle blocks and turned texts, and its own renderer draws them.
Plans that keep their place, Media Studio touches, and Export chat
Desktop app
A plan stays with the chat while you answer the model's questions, so the step you answered is ticked and "Continue the plan" goes on instead of starting over.
A trip abroad no longer gets the timetable offer, which covers Norway only.
When a model says it is using a tool but doesn't, GlirOS asks it once to do it or say it can't.
"Next Monday" and "Friday" are given to the model as dates, and made-up source lines are no longer shown.
Export chat (Settings > Chat history): adds Export chat to each chat's ⋯ menu, a file for troubleshooting. Keys and passwords are left out.
Media Studio: moving a file to the timeline with changes not made yet asks first, and can make them before it moves.
Media Studio: crop a clip right on the timeline, even when it is the only one.
Media Studio: move the red line with a bar under the preview, one frame at a time, or by dragging the scissors.
Tools
Travel planner is now called Timetables (live departures only), and the Trip planner plans the trip, with an outline when something has no live source.
Media gallery, Translator, Sports lookup, Picture to 3D, SCORM and the Trip planner can tell you when a long job is done.
When GlirOS can't get a fact (a flight abroad, a hotel, a price), the model says so and still gives you the outline, with placeholders, instead of making things up.
Switching a tool on or off in Settings while you chat is now known to the model in that chat.
One decision at a time: "Start a project for this" waits until you've answered the plan's tool question.
Small models no longer show GlirOS's own notes ("[Focus: …]", "[GlirOS log …]") in their answers, and a very short message like "great" is answered in the chat's language.
A tool switched back on gets its "Open" buttons back in the chat right away.
Thorough answer: for questions about law, health, safety, money or immigration, GlirOS looks at official sources first, and shows emergency numbers before the answer.
Compare models (Settings > Models): ask the same question of up to three models and keep the answer you like.