batty-251/voxai-coder-releases

By Batty 251 (@batty-251) · Licence: MIT · 87 published versions · 3 files · 24 KB of source · Last published 2026-10-18

A public project on CodeRook. Browse it at https://coderook.com/batty-251/voxai-coder-releases.

Versions

• v0.2.86 (2026-10-18): 0.2.86: Several projects at once in one window: going back to a project whose run is still going no longer breaks it, and Stop ends a run waiting on an approval. The balance counts down as runs spend. Side by side, the session list opens in — Several projects at once in one window: going back to a project whose run is still going no longer breaks it, and Stop ends a run waiting on an approval. The balance counts down as runs spend. Side by side, the session list opens into the pane you are working in. Releases now come from the main VoxAI Coder project.
• v0.2.85 (2026-10-17): 0.2.85: Spends less on every model: one fixed prompt per session so the provider cache keeps hitting, only the memory facts that matter, the thinking level shown with its cost, long requests kept under GPT and Grok price jumps, stale reads — Spends less on every model: one fixed prompt per session so the provider cache keeps hitting, only the memory facts that matter, the thinking level shown with its cost, long requests kept under GPT and Grok price jumps, stale reads and old screenshots left out, and a stop when the model goes in circles. Edits that match more than one place are refused. The model picker shows what a message costs you. Updates download with progress in Settings > Account.
• v0.2.84 (2026-10-17): 0.2.84: Settings redone page by page: saved keys shown as themselves, sortable usage, Add credits, model pickers by company, a picture of the context, Memory with Undo, and Skills grouped with New skill. Local models: download GGUF models f — Settings redone page by page: saved keys shown as themselves, sortable usage, Add credits, model pickers by company, a picture of the context, Memory with Undo, and Skills grouped with New skill. Local models: download GGUF models from Hugging Face, checked against their checksums, and see which fit your card. A new Apps (MCP) page to add and test MCP servers. Finished turns fold to their answer, answers stream formatted, queued messages send after a Stop, long sessions keep what you asked, and New update ready can be dismissed.
• v0.2.83 (2026-10-04): 0.2.83: Batch models are no longer offered: OpenRouter answers them only within 24 hours, so they failed in chat. Models with a flex tier (OpenAI, Google) get a Cheaper, slower switch in the model picker: the batch price, live. Replies are — Batch models are no longer offered: OpenRouter answers them only within 24 hours, so they failed in chat. Models with a flex tier (OpenAI, Google) get a Cheaper, slower switch in the model picker: the batch price, live. Replies are slower; when flex is busy the normal price applies instead of an error. Sessions saved with a batch model move to the same model on flex.
• v0.2.82 (2026-10-03): 0.2.82: New: Review changes (the ± button) to see the diff hunk by hunk, put changes back, comment on lines for the agent, write the commit message, commit, push, open a pull request and fix failing checks. Type @ to send files with your me — New: Review changes (the ± button) to see the diff hunk by hunk, put changes back, comment on lines for the agent, write the commit message, commit, push, open a pull request and fix failing checks. Type @ to send files with your message. A terminal panel (>_) beside the chat. Fixes: queued messages go to the right session, safer shell approvals, sessions that cannot be corrupted, Stop works while a local model loads.
• v0.2.81 (2026-10-03): 0.2.81: Steering works mid-run: a message you steer in while VoxAI is working is read at its next step, shows as a Steering conversation bubble, and leaves the queue. Queued messages still send if a run's end was missed. Linux: each release — Steering works mid-run: a message you steer in while VoxAI is working is read at its next step, shows as a Steering conversation bubble, and leaves the queue. Queued messages still send if a run's end was missed. Linux: each release gets a VoxAI-Coder-<version>-Linux.sh installer next to the AppImage.
• v0.2.80 (2026-09-28): 0.2.80: Screenshots now work while VoxAI Coder is minimised, so it can still check what it builds. The model picker shows how fast each local model runs on your PC and suggests the biggest one as the Team coder. Local models no longer fail — Screenshots now work while VoxAI Coder is minimised, so it can still check what it builds. The model picker shows how fast each local model runs on your PC and suggests the biggest one as the Team coder. Local models no longer fail when a long conversation outgrows their memory; older turns are summarised and the request is sent again.
• v0.2.79 (2026-09-28): 0.2.79: Local models: bigger mixture-of-experts coders such as Qwen3-Coder-30B-A3B now load well on an 8 GB card, with part of the model in RAM, and are tuned for your PC. Local Team coders get 30 minutes a step. The slow-PC warning no long — Local models: bigger mixture-of-experts coders such as Qwen3-Coder-30B-A3B now load well on an 8 GB card, with part of the model in RAM, and are tuned for your PC. Local Team coders get 30 minutes a step. The slow-PC warning no longer fires by mistake while a prompt is read, and only a truly starved local coder is stopped. The graphics card is found reliably at startup.
• v0.2.78 (2026-09-28): 0.2.78: Local models: if another program slows the PC so much that the local model crawls, VoxAI now says so within half a minute instead of waiting silently. A Team coder on a crawling local model stops after 2 minutes and the planner tell — Local models: if another program slows the PC so much that the local model crawls, VoxAI now says so within half a minute instead of waiting silently. A Team coder on a crawling local model stops after 2 minutes and the planner tells you why, instead of trying again for half an hour.
• v0.2.77 (2026-09-28): 0.2.77: Local models: run GGUF models on your own PC for free. Settings > Local models installs a llama.cpp engine, finds your GPU and tunes how much of the model it holds, and remembers it. Local models appear in the picker at $0 and can b — Local models: run GGUF models on your own PC for free. Settings > Local models installs a llama.cpp engine, finds your GPU and tunes how much of the model it holds, and remembers it. Local models appear in the picker at $0 and can be the Team coder. They save their work to files instead of pasting code in the chat.
• v0.2.76 (2026-09-28): 0.2.76: Auto: new safety rules stop and ask before deploys, cloud changes, remote databases, ssh, running downloaded programs or scripts from outside the project, wiping drives, and changing machine-wide tool settings. Settings > Account ca — Auto: new safety rules stop and ask before deploys, cloud changes, remote databases, ssh, running downloaded programs or scripts from outside the project, wiping drives, and changing machine-wide tool settings. Settings > Account can now list every key on your OpenRouter account and what each has spent. Updates keep a log, and Restart says when it can't go ahead. Text that arrived while the window was covered shows at once when you come back. Team planners spend fewer turns on bookkeeping.
• v0.2.75 (2026-09-27): 0.2.75: Team: the planner hands out small steps and keeps working with the same coder instead of starting a fresh one each time, so the coder keeps what it has read. The coder asks when it is stuck instead of guessing, and tests with a shor — Team: the planner hands out small steps and keeps working with the same coder instead of starting a fresh one each time, so the coder keeps what it has read. The coder asks when it is stuck instead of guessing, and tests with a short run-through.
• v0.2.74 (2026-09-27): 0.2.74: Team: the planner hands off, trusts the coder's report and takes one final look, so builds cost less; if the coder can't see images, the planner does the looking. Coders now see their own screenshots. Auto: the main model decides an — Team: the planner hands off, trusts the coder's report and takes one final look, so builds cost less; if the coder can't see images, the planner does the looking. Coders now see their own screenshots. Auto: the main model decides anything the risk rules don't flag, with no extra checker call. A session moved back from side-by-side shows its full history. A model request that stops responding is retried after 3 minutes instead of waiting forever.
• v0.2.73 (2026-09-27): 0.2.73: Team: the coder now has the browser. It opens and uses what it builds (clicks, plays) and fixes what looks or plays wrong before reporting; if it hands back a page or game it never looked at, the planner sends it back. — Team: the coder now has the browser. It opens and uses what it builds (clicks, plays) and fixes what looks or plays wrong before reporting; if it hands back a page or game it never looked at, the planner sends it back.
• v0.2.72 (2026-09-27): 0.2.72: Models OpenRouter won't run in VoxAI are greyed out in the model picker with the reason. Recent models leaves out ones that can't do the job. Pop-up messages sit over the chat instead of under the preview. Team: when the planner kee — Models OpenRouter won't run in VoxAI are greyed out in the model picker with the reason. Recent models leaves out ones that can't do the job. Pop-up messages sit over the chat instead of under the preview. Team: when the planner keeps trying to do the coder's work, it is told more plainly to hand it off.
• v0.2.71 (2026-09-27): 0.2.71: Team: the planner hands the writing and the browser testing to the cheaper coder, so Team costs less. Free models: when one is busy or out of its daily allowance you get one-click Try again, the paid version, or another free model; — Team: the planner hands the writing and the browser testing to the cheaper coder, so Team costs less. Free models: when one is busy or out of its daily allowance you get one-click Try again, the paid version, or another free model; today's free requests show beside your balance. Browsers left open in the background stop using the CPU when nothing is using them. Dialogs and menus are no longer covered by the preview or side-by-side sessions. The preview now shows a new session's page straight away.
• v0.2.70 (2026-09-27): 0.2.70: Free models: a busy free model is no longer reported as out of its daily limit, free models get longer retries, a refusal of one model is no longer called a bad API key, and more free providers accept VoxAI's requests. — Free models: a busy free model is no longer reported as out of its daily limit, free models get longer retries, a refusal of one model is no longer called a bad API key, and more free providers accept VoxAI's requests.
• v0.2.69 (2026-09-27): 0.2.69: Agent browsers keep running while out of sight: a game being tested in a session's browser no longer freezes when that browser isn't the one shown in the preview. — Agent browsers keep running while out of sight: a game being tested in a session's browser no longer freezes when that browser isn't the one shown in the preview.
• v0.2.68 (2026-09-27): 0.2.68: Each session now has its own browser, so agents working side by side no longer fight over one page, and the preview has a tab for each session's page. While a planner waits on its coders, you see what they are doing. Runs started in — Each session now has its own browser, so agents working side by side no longer fight over one page, and the preview has a tab for each session's page. While a planner waits on its coders, you see what they are doing. Runs started in a split pane show as running everywhere.
• v0.2.67 (2026-09-27): 0.2.67: Fixed empty 'Try it / Not now' buttons showing during runs. All Grok models are now in one group in the model picker (OpenRouter labels them SpaceXAI and xAI). Every request starts about 1,100 tokens smaller. Auto now always asks be — Fixed empty 'Try it / Not now' buttons showing during runs. All Grok models are now in one group in the model picker (OpenRouter labels them SpaceXAI and xAI). Every request starts about 1,100 tokens smaller. Auto now always asks before clicks or typing that buy, post, sign in, delete accounts or data, uninstall, accept terms or type a password.

Files in the latest version

• CHANGELOG.md
• LICENSE
• README.md

README