Changelog
Shipped, fixed, and improved
Every 1LocalAITool release, straight from the release notes that ship with the product.
SHIPPED · v0.13.2
1LocalAITool v0.13.2 — compared answers named on the phone, and Send on the first tap
Improved
- A compared answer says which model wrote it. On the phone page the two answers to one message sat one under the other with nothing to tell them apart. Each now carries its model's name, first model first, and the bar at the top reads *Qwen 2.5 1.5B vs Gemma 2 2B*. While the turn runs, the first model's answer is written under its name and the second model's name waits below it.
Fixed
- Send takes the first tap. On some phones the first tap on Send only put the keyboard away, and it took a second to send. One tap sends now.
Fixed
- Lists in an answer have their bullets and numbers back. A list the model wrote showed as indented lines with nothing in front of them. Bulleted lists now show a dot, a list inside a list a ring, numbered lists their numbers, and a checklist its boxes — in chat, in a job's result and in the artifact panel.
SHIPPED · v0.13.0
1LocalAITool v0.13.0 — two models side by side, chats dragged into folders, and a whole chat saved as a file
New
- Drag a chat onto a folder. Press a chat in the list and drag it to a folder; the folder lights up and the chat lands in it. Drag it onto *Chats* to take it back out. Esc cancels.
- Pick several and file them together. Ctrl-click (⌘-click on a Mac) adds chats one at a time, Shift-click takes everything in between. Drag any one of them and they all go — or use *Move to…* in the bar that appears, and *Move N chats to folder* in a chat's own menu, which need no dragging at all. Filing leaves the order of the list alone.
- Copy chat takes the whole conversation. One button in the chat's header copies everything since the first message, each part under the name of who wrote it.
Improved
- Every answer has its own Copy. The buttons under an answer used to appear only under the pointer, which read as "only the last answer can be copied". They are always there now, under every answer and every message of yours. *Copy plain* beside it gives the same words without the formatting marks.
- Each answer says which model wrote it. Answers from this release on keep the name of their own model on the line beneath them, in a comparison and out of one, so changing a chat's model halfway no longer puts the new name under what the old one wrote.
Fixed
- Yellow when paused, red when stopped. The lamp in the sidebar stayed green while the app was paused. It is yellow now, and so is the *Pause* switch while it is on — here and on *This computer* — so the two agree. An engine that has given up shows red and reads *Stopped* until *Try again*.
SHIPPED · v0.12.0
1LocalAITool v0.12.0 — tools from other apps, and an Updates tab that understands a subscription
New
- Add another app's MCP server under Skills. Under *Tools from other apps*, name the app and paste the MCP address it shows — a FluentBoards or FluentBooking site, a CRM, any MCP server over HTTP. Its tools join every chat here, and any job whose chip you tick on the Tasks screen. Each call goes to that app and nowhere else, and the strip under the answer says what happened: *asked FluentBoards to list boards*.
- Sign in the way the app asks. A WordPress username with an application password, a bearer token, or a header of the app's own. The app builds the Authorization header itself — nothing to base64 by hand. The password is encrypted on this computer, kept apart from the settings, and never shown again; editing with the field left blank keeps it.
- Test before you save. *Test* lists the tools the server offers, or says plainly why it could not: a refused sign-in, a wrong address, a page where an endpoint should be. *Check again* asks a saved app for its tools afresh, and the switch on each row turns its tools off without forgetting it.
Improved
- The Developer screen's MCP tab says which way is which. It was only ever the snippet for an agent to use *this* computer, and read as if it were the place to set up the other direction. It now says so, lists the read_pictures tool it had left out, and points at Skills.
Fixed
- A lapsed All Access subscription is said plainly. The Updates tab used to tell a subscriber that a newer build was past their update window, which is a lifetime licence's story. It now says the subscription lapsed, with the date it was paid through, and *Renew subscription* opens the store's key page. The Free-versus-Pro table reads *Included while you subscribe* on its updates row, and the upgrade prompt stops selling a lifetime to someone who already subscribes.
- The Complete bundle's current prices. The bundle card now shows $199 for one device and $349 for three, as the store does.
SHIPPED · v0.11.0
1LocalAITool v0.11.0 — the graphics card used and accounted for, your own picture checkpoints, and StoicSoft bundle keys
New
- Running on the processor is never silent. When there is a card and text, pictures or 3D end up on the processor, a dialog says so, with the reason in the engine's own words, *Logs and a report*, and *Try the card again*. The sidebar light is green on the card and amber off it. This also covers a text engine that starts but finds no card it can use — a missing Vulkan driver used to leave it on the processor with the light still green.
- Graphics card only. A new switch under *Logs and crash reports*, the opposite of *Use the processor only*: when the card cannot do something, text, pictures and 3D stop and say why instead of moving to the processor.
- This computer says what to expect before anything runs. One line per feature — text, pictures and video, 3D — says whether it runs on the card or the processor on this machine, and why.
Improved
- Pictures and 3D use an NVIDIA card on Windows. Pictures fetch the CUDA build with its runtime, then Vulkan, then the processor, so an RTX card no longer draws on the processor. 3D installs PyTorch built for the card. A driver older than 525 skips CUDA rather than failing on the first picture.
- Text uses an NVIDIA card on Linux. The engine now fetches the Vulkan build, which every NVIDIA and AMD driver runs. An engine installed before is replaced the next time the app starts, before anything can start the old one.
- Another program on the app's port is named. A banner says whether it is another copy of the app or some other program, instead of retrying in silence.
Fixed
- macOS 15 makes pictures and 3D models again, without being told to upgrade to macOS 26. The newest picture build made for macOS 15 is found further back in the release list. A known-good build is kept as a last resort, so this keeps working as newer releases push it further back, and when GitHub limits how often the list can be read.
- Try the card again really tries again. It restarts text on the card, fetches the card's picture build afresh with the next picture, and reinstalls the card's PyTorch for 3D when that was the cause. A fallback is no longer kept for good.
- The speed test ends. It is capped at 96 tokens and three minutes, and reads the engine's own speed figure. It could sit on *Measuring this computer* for as long as the model kept writing.
SHIPPED · v0.9.0
1LocalAITool v0.9.0 — a window that follows the system into dark mode, and an engine that says why it will not start
New
- The window follows the system's light or dark setting. A new theme, *Match this computer*, wears Paper by day and Night shift once macOS, Windows or the Linux desktop goes dark, and changes the moment the system does. It is the default on a new install; a theme already chosen stays as it was.
- Light or dark in one click. A switch at the bottom of the sidebar steps through *Match this computer*, Light and Dark, without a trip to *This computer*.
- The first download can wait. The last setup step has *Skip the download*: nothing is fetched, and the app opens on *This computer › Models*, where a model you already have can be added, or one installed when you are ready.
Improved
- Download progress stays in sight. While a model downloads or moves, the sidebar shows how far along it is on every screen, and the model you pressed *Install* on carries its own progress bar.
- Install is green, Remove is red. On *This computer › Models* the two read apart at a glance, and a pressed button stays pressed — *Removing…* — until the work is done, instead of springing back after a moment.
- Moving the models no longer means typing a path. Under *Keep them somewhere else*, *Browse…* picks the folder and connected drives are one-click choices. The section says where the models are kept now, and after a move, where they went. Read-only drives, like a mounted installer, are not offered.
Fixed
- A message that gets no answer is not counted. On the Free plan a chat message counts when its answer arrives, not when it is sent. A turn the engine never answered, or one stopped before it said anything, costs nothing.
- The graphics card gets what fits, not everything. The engine was told to put the whole model on the card, sized against the computer's memory; on a card with less memory than the computer, that failed at every start. It now loads as much as the card holds.
- GPU compositing in the window is off. WebKit's GPU rendering is the likely cause of a toolbar that flickers and keystrokes that go missing while the engine works the card, so it is switched off on Linux. ONELOCALAITOOL_GPU_UI=1 turns it back on.
SHIPPED · v0.8.0
1LocalAITool v0.8.0 — a Linux build that uses your graphics card, and models that stay put
New
- A Linux build, next to the Mac and Windows ones. Each release now carries an AppImage and a .deb for 64-bit Intel and AMD machines, and both update themselves the way the other builds do. They are built on Ubuntu 22.04, so they run there and on anything newer. On Ubuntu 24.04 and later the .deb is the simpler install: an AppImage there needs the libfuse2t64 package first.
- Making pictures on Linux needs Ubuntu 24.04 or a distribution as new. The picture worker is published built there, and an older system cannot start it.
Improved
- An NVIDIA card on Linux is used. It used to fall through to the CPU build of the worker; NVIDIA and AMD cards now get the Vulkan build, for chat and for pictures, which runs on any driver that ships Vulkan.
Fixed
- Installed models no longer rename themselves at every launch. 0.7.0's launch-time look through models/ took each installed model for a file dropped in by hand: it was added again as a custom model under a new name — moved out from under the engine that was loading it — and the real model, now looking missing, started downloading a second time. Files the app keeps under their own name are left alone now, and the first launch of this version puts back what the bug did: the file returns to its model, the extra entries leave *This computer › Models*, and the second download is dropped.
- A models folder on an unplugged drive is said so plainly. With *Kept in* on an external drive that is not connected, the app no longer makes the folder again — on an empty mount point that quietly started a second, empty library on the internal disk. *This computer › Models* says the folder is not connected, downloads wait for it instead of failing with a path error, and plugging the drive back in is enough: the next run starts from it without a restart.
- A busy GitHub no longer reads as “error decoding response body”. When GitHub turns away a worker download because too many came from one address, setup says so and how long to wait — it used to report a garbled body, or a pack “missing its build for this computer”.
SHIPPED · v0.7.0
1LocalAITool v0.7.0 — one plain copy of each model, and files that say what they are
Improved
- No more loaded/. Text models used to be stored encrypted as .gguf.enc and decrypted into loaded/ every time the engine ran — a second multi-gigabyte copy for nothing it could not do by reading the file itself. Models now live as plain .gguf in models/ and the engine reads them there, like every other tool. A copy left in loaded/ by the old scheme becomes the library copy once; an orphaned .gguf.enc is decrypted in place the first time the model runs, and never again. Chats and jobs stay encrypted — that was always the part worth encrypting.
- Files are asked what they are. A .gguf is no longer assumed to be a chat model: the app reads its architecture and a .safetensors file's tensor names, so a diffusion checkpoint — FLUX, SDXL, z-image — files itself under the picture models, and an encoder or a language-model safetensors says so plainly instead of listing as a model that cannot run. Links get the same answer from a small range read, before any download.
- Anything dropped into models/ is found at launch, not only when Re-scan is pressed — copy a checkpoint in and it appears in the pickers.
SHIPPED · v0.6.0
1LocalAITool v0.6.0 — your own models, kept tidy, and an audio cleaner that fetches itself
New
- .safetensors is welcome. Links and files ending in .safetensors now join the picture models instead of being turned away — a .gguf lists with the text models, a .safetensors with the checkpoints, and either can be chosen under *Which model* on the Tasks screen when a picture or a 3D model is being made.
- Files copied into models/ by hand are found. A *Re-scan* button beside *Kept in* looks through the folder again: a file named like a catalog model becomes that model, an unknown .gguf joins your added models, a .safetensors joins the picture checkpoints.
- A "Locate file" button finds a model for you. With the path box empty, *Add file* becomes *Locate file* and opens a file picker already filtered to .gguf and .safetensors; a typed path still adds directly.
Improved
- Every added model is kept only once, encrypted. A .gguf brought in — by file, by Install, by Re-scan — is encrypted to .gguf.enc and the plain original is deleted; a plain duplicate left next to its encrypted copy is cleared away too.
- Cleaning audio no longer asks you to install ffmpeg. The audio cleaner is fetched into the app's own folder the first time a *Clean audio* job needs it — with progress in the setup card like every other worker — and listening jobs borrow it only when a recording's format needs converting first.
Fixed
- A bad link can no longer look like a model. A link that answers with an error page used to be added anyway and sit in the list as a mysterious *0 B* model that never downloaded. The app now asks the link what it would return and refuses a 404 or anything too small to be a model file.
- loaded/ can no longer collect a second copy of a model. The decrypted copy the engine reads was still being left behind by a failed start or a crash partway through — llama31-8b among them. loaded/ is now swept down to the one file the engine is running or bringing up, whenever the queue ticks — a stray dropped in by hand disappears within seconds, and an engine crash clears its leftovers the same way.
SHIPPED · v0.5.0
1LocalAITool v0.5.0 — pictures that draw on Windows, and a model folder that stops doubling
Fixed
- Picture models failed to start on Windows. The drawing worker and the chat worker unpacked into one shared folder, and both carry identically named support files — whichever arrived last overwrote the other's, which Windows reported as sd-cli.exe — Entry Point Not Found on every picture. The drawing worker now unpacks into a folder of its own, checks it can actually start before calling itself ready, and clears away the leftovers of the mixed install so a working set is what it finds next time.
- The model folder no longer stores every model twice. Each model is kept encrypted; the copy the engine reads while it runs used to stay on disk forever, so a second full copy of every model ever opened piled up — and moved along on a *Kept in* change. The working copy now leaves with the engine, strays are swept when the app starts and before each run, and a move carries only the real files. After a move, the new home holds models alone; the working copy is remade only while a model is in use.
New
- Add a model by link. Paste a link to a .gguf file — on Hugging Face it is the address behind a file's download button — and it joins the list under *This computer › Models* with an Install button like any other. It never nominates itself: it waits for Install, then Use this. If the host will not say how big the file is, the card says so and learns it when the download lands.
- Add a model another app already downloaded. *Look for them* finds the .gguf files LM Studio, the Hugging Face cache and Downloads already hold — or paste a path yourself. Adding one needs no new download: it is brought into the app's encrypted home and runs exactly like a fetched model.
- Edit & run again sits beside Run again on a failed, cancelled or finished job. The ask comes back prefilled — change the title or the text, save, and it queues again from scratch. The phone inbox offers the same.
Fixed
- Remove works for every added model. It unlists the card and deletes its files, whether or not the model was ever installed.
New
- Rename the computer without starting over. *This computer › Overview* gained a *Call this computer* field; the new name shows up on the pool, the queue and the phone at once.
- Local server, like the one in LM Studio. The Developer screen has a new *Local server* card with every address this computer answers on — this computer only, this whole network, Tailscale and Anywhere — each one a tap to copy, next to the snippets for Claude Code, Cursor, Terminal, curl, Python and MCP.
- A public address with Cloudflare Tunnel. One switch and the app fetches cloudflared itself, brings up a https://…trycloudflare.com address anyone can reach, keeps it across restarts, and pairs phones with it so cellular works too. Callers arriving through the tunnel still need an API key — the address alone is not one.
Improved
- Progress you can watch while things install. The installer used to show only the model's progress; everything else looked like a stall. A bar and a note now cover every step — the worker, the model, the checks and the engine warm-up — with how much has arrived and how fast.
Fixed
- The setup wizard froze on the last step. The last step deadlocked on the work queue itself, so it never finished — though setup had already been saved, which is why closing and reopening landed inside the app. The step now finishes and hands off properly.
- "Worker binary missing from pack" on Windows. The download picked the wrong file from the list — a small helper pack instead of the real worker — so installs stopped before anything ran. The right pack downloads now, along with the extra files the Windows NVIDIA build needs to start at all.
2 older 1LocalAITool releases are listed in the release history.












