Changelog
What's new in Consoles
Every feature, improvement and fix, release by release.
Download for Mac
Latest: v0.8.10, November 17, 2026
Activity1,911 commits in the last six months
MayJunJulAugSepOctNov
LessMoreRelease
New GPT models and steadier compaction
Changes
- New GPT models — ChatGPT and OpenAI connections get the latest GPT models with up-to-date pricing, and requests to OpenAI's Responses API work with its newest changes.
- Pipelines setup — A pipeline's runner now downloads the project code from GitHub when it can, with a fallback, so setting up a runner is faster and more reliable.
Bug Fixes
- Context meter after compaction — The context meter resets when a chat is compacted, and the different ways a chat can be compacted now behave the same.
No more stuck "Compacting…" and a crash fix
Bug Fixes
- "Compacting conversation…" no longer stays up — After Claude compacted a long chat, the "Compacting conversation…" line could stay at the bottom of the chat until the whole turn finished. It now clears as soon as compaction is done. If a compaction fails, the line clears and a short note says it didn't finish, and a compaction that never finished no longer takes over the marker of the next one.
- Crash from a noisy browser tab — A page in the in-app browser stuck in an error loop (for example a dev server page after its server stopped) could log thousands of errors a second and crash Consoles. Consoles now logs at most a few of those lines a second per tab.
Long source calls that come back on their own
Changes
- Built-in Video source hidden — The built-in Video (Rend) source failed for everyone signed in, showing "Video is temporarily unavailable", and chats were still being sent to it for video work. It is now hidden and chats are no longer pointed at it. A Rend source you connect with your own account works as before.
- Long source calls return right away — Source connections now use the newer MCP client, which still connects to older servers. When a source answers a long call with a background task, the call returns at once, Consoles follows the task, and the chat is woken with the final result when it is ready.
- Advisor for agents — In Settings → Agents, an agent can now have an Advisor (Fable, Opus or Sonnet). On Claude, the agent's own model does the work and checks with the advisor at key decisions. Off by default.
- Pipelines — A new Pipelines panel with runner settings (secrets, cloud switch, push hook, delete), a page across all projects, a visible status line in the chat, a run banner, and a dot on projects whose runs failed.
- Commit from a chat — Chats have a commit button for the changes they made.
- Claude Code updated — The bundled Claude Code is now 2.1.325.
Bug Fixes
- Messages stuck behind background work — A message sent to a chat with background work running could wait in the queue until that work finished, sometimes for an hour. Messages now go in right away; if one still can't get through, the background work is stopped after 10 seconds and the agent is told what was stopped.
- Tool search loops on Claude chats — On Claude chats, tool search kept telling the agent to repeat the search for a source's tools it could never load, so helpers retried over and over and gave up. It now connects the source and tells the agent how to load the tools.
- Finished answers stay visible — When a turn's work is folded away, its finished answer now stays visible.
No more crashes on very long chats
Bug Fixes
- Crash at the end of a turn — In a very long Claude chat, the check that keeps saved keys and protected values out of the chat's history could use up all of the app's memory when a turn ended, and Consoles would quit. It now works through the history in small pieces, so memory stays flat however long the chat is.
Compact no longer gets stuck
Bug Fixes
- Compact finishes — Running /compact right after a turn that had background work (a helper or a background command) could leave the chat stuck on "Compacting", with every later message waiting in the queue. The compaction itself had finished; the chat just never heard about it. It now ends normally and your next message runs.
Chats keep to their own files
Improvements
- Your changes, not everyone's — When several chats share a folder, each one now only claims the files it wrote with its own edit tools (and those of its helpers). Shell commands like git stash, git pull or a long type check no longer pull other chats' uncommitted work into its file list, turn card or Undo.
- Helper questions reach you — When a hidden helper asks a question, it now shows in the parent chat, and your answer goes back to the helper.
- Fewer refused tools — Helpers are no longer offered tools they aren't allowed to use, such as starting their own helpers or anything their agent profile turns off.
For advanced and headless setups
- Orchestra can be customized per workspace: orchestra/consigliere.md and captain.md replace the built-in instructions, and orchestra.json adds rules (allowed tools and helpers, or a block with a message). Switching Orchestra on can keep the current model, and the server status lists these features.
- An agent marked as able to delegate can start one level of helpers from its own lane; their progress shows in the main chat, and stopping the agent stops them too.
- Headless hosts can drop unused built-in Claude tools with the CONSOLES_DISALLOWED_TOOLS setting.
Venice replaces NEAR AI
Improvements
- Venice only — NEAR AI is gone from the connection presets; Venice now serves DeepSeek V4.1 Flash 1017. Chats that used NEAR's DeepSeek models carry on with Venice's matching models.
- Images on OpenAI-compatible endpoints — Images are on by default, so Venice's vision models get your screenshots without a text-only warning. Known text-only models stay off.
- Pricing — Every Venice request is priced at Venice's rates: new input, cached input, output (thinking included) and images, streamed or not.
Venice
Features
- Venice — Add your Venice key under Settings → AI to run DeepSeek V4.1 Flash 1017, uncensored or regular, with image input, plus Venice Uncensored. Every effort level works, prompt caching is on, Venice's own system prompt is turned off so Consoles' instructions apply, and every request is priced at Venice's rates.
Bug Fixes
- Model divider — A reply that is still streaming no longer shows an "unknown" model-switch line above it.
- Windows — The installer builds again.
Full chat cost, readable thinking, Windows build
Improvements
- Chat cost — Click the context ring to see what the chat has cost, including its subagents and compactions. When there's something to show, it also lists the cost since the last compaction and the subagents' share. The cost no longer sits next to the ring.
- Thinking — The live thinking line shows plain text (no raw markdown), changes only when a sentence finishes, and opens to the full reasoning with a click.
Bug Fixes
- Edited files card — A chat that ran no tools no longer claims edits made by another chat in the same folder.
- Windows — The installer build compiles every bundled server again.
NEAR AI, Linux and the cost of every chat
Features
- NEAR AI — Add your NEAR AI key under Settings → AI and run DeepSeek V4.1 Flash on NEAR's private servers, including the abliterated snapshot for security research. Effort levels map to what NEAR accepts (Off really turns thinking off), replies can run to 64K tokens, prompt caching cuts repeated input by about 98%, and every request is priced.
- What a chat costs — The composer shows the chat's running cost next to the context ring on pay-per-use connections, and the context popover lists it on every connection (as the API price on subscriptions). Every request and tool call counts.
- Linux — Consoles is available for Ubuntu and other x64 Linux as an AppImage. consoles.ai/download picks it on Linux, and
curl -fsSL https://consoles.ai/install | bashinstalls it.
Improvements
- Model picker — OpenAI-compatible endpoints show under their own name (NEAR AI, Manifest, OpenAI Compatible) with their own models, and GPT-6.1 Terra sits after Astra, in Codex's order.
- Readable thinking — The live thinking line skips code and shows the last finished sentence of long reasoning, so raw-reasoning models like DeepSeek no longer flicker.
- Helper chats — Each connection can set its own early-compaction size for helper chats.
Bug Fixes
- Editing a connection — Saving without retyping the key no longer tests the masked placeholder, and branded endpoints reopen under their own name.
- Windows builds — The installer build no longer runs out of memory.
Org controls, PII protection and Windows parity
Features
- Org controls — Organizations get teams, model rules for the org, a team or a person (allowed providers and models, enforced in the app through a signed policy), an audit log with CSV export, and AI usage insights on
/org: spend, tokens, top models and active people, from org to team to person. Only daily totals leave the device; prompts, chat titles and files never do. - PII protection — Emails, phone numbers, cards, SSNs and IBANs become placeholders like
devivo@gmail.combefore the model sees them. The real values live in an encrypted vault for each workspace and are filled back in when a tool runs. Reveal chips, a Privacy section in Workspace settings, and roster seeding for known names. Cursor is blocked while protection is on. - Org library — Share skills, sources and agents with your organization, and they sync into everyone's workspace.
- Long tasks in the background — Any model can move a long-running command to the background and keep working;
watch_mcp_jobwaits on long source jobs and wakes the chat once they finish. - New models — Claude Fable 5.2 and GPT-6.1 Terra.
- Local voice input — Dictate with on-device transcription.
- 3D previews — STL and 3MF models render inline in chat.
- Crash reports — Opt-out crash reports and usage stats (Settings → App) help us fix bugs. They never include your chats, files or keys.
Improvements
- Clearer first launch — Use Consoles free on this computer or sign in, as two equal choices. The sidebar shows "Free · This computer" or your organization, and Cloud computers are marked as coming soon.
- Windows — Taskbar badge for unread chats, a flashing taskbar button when an agent is waiting, Ctrl+= / Ctrl+- / Ctrl+0 zoom, Check for Updates in the menu, Show in Explorer, Windows privacy steps for the mic and camera, more Open-with apps (JetBrains, Sublime, Zed, Notepad++, PowerShell), and PowerShell as the terminal when Git Bash isn't installed. File names no longer show whole
C:\paths. - Faster, lighter chats — Chat files stay lean (no duplicated images or icons), steering no longer rewrites big session files, the sidebar and tab strip stop re-rendering on every update, and idle browser tabs off screen release their page.
- Steadier helpers — Background helpers survive turn handoffs and errors, messages sent while they run join the live stream instead of interrupting, and Steer on a queued message sends it right away when nothing is running.
- Commit chip — Commits only a run's own changes and flags leftover hunks.
- Web search — Parallel search with a fallback provider.
- Hosted agent profiles and native Responses endpoints for OpenAI-compatible hosts.
- Sign-in — The browser says you're signed in only after the app saved it, Windows brings the app to the front afterwards, and a clear message replaces raw credential-store errors.
- Safer workspace secrets — Commands resolve only the keys they reference, with
--env NAMEfor environment access; temporarysecret run --injectis gone (use env or stdin, or opt into permanent injection with its warning).
Bug Fixes
- Usage stats no longer freeze the app when a window gets focus.
- PII stops treating filing and trade dates as birth dates.
- Model switches no longer truncate history, and orphaned tool results no longer break OpenAI Responses chats.
- Claude sign-in refreshes when under 2 hours are left, so long chats keep working.
- Command and download hardening — tighter Plan-mode Git and GitHub rules, corrected macOS admin-command quoting, and SHA-256 checks before installing downloaded cloudflared.
Breaking Changes
secret run --injectis removed. Use--env,--env-fileor--stdin.
Secrets, schedules and a voice for your workspace
Features
- Workspace secrets — Save keys in an encrypted, write-only workspace vault (Settings → Secrets) and reference them as named lock chips in chats. Claude, Pi, and Cursor can use placeholders through source tools and the bundled
secretcommand without putting raw values into chat instructions; known values are masked in command output, app transcripts, logs, live publishing, and reports. Claude and Pi also mask model-facing tool results. Permanent file injection still writes real keys to disk and is not a security sandbox. - Repeat a chat on a schedule — Repeat… on any chat sets a schedule that runs in that same chat and project. A tray in the composer lists the schedule and jumps to each run, and runs missed while the app was closed catch up. Model and thinking level carry into every run.
- Voice concierge — Talk to your whole workspace from the voice button beside Go live. It reads your sessions live, answers "what did that chat say", asks captains and subagents for you, and takes actions with a confirm card.
- Sign in with Cursor and SuperGrok — Run chats on your Cursor plan or your SuperGrok / X Premium plan, with Grok 4.7, native Cursor compaction, steering and resume. Grok's X Search shows as its own tool with the tweet links, and SuperGrok credits show in Usage.
- Speed gauges — A little speedometer in the composer shows this chat's live tokens per second (its spawned chats and captains included), and one in the top right shows all of Consoles. The Usage page and Home add tokens per second, agents at once and agents per day.
- Steer instead of stop — Sending mid-run steers the agent after its current step instead of interrupting it. The steer shows in the chat right away with Undo until it's read, and you can still pick queue or interrupt.
- Pull requests and issues in the right panel — New Pull requests and Issues tabs, an assigned-issues widget on Home, and PR chips on session rows that roll captains' PRs up to their parent chat and flag new review feedback.
- Commit chip — Each chat shows a commit chip with uncommitted, committed and pushed states across every repo it touched, a repo picker, progress while it runs, a remembered push mode, and Fix with AI when a pull or push fails.
- More on your live stream — Viewers see your right panel's browser (as video, privacy-guarded) and terminal, a HUD with the chat, lines changed, commits and streak, reactions, predictions with votes, points and a leaderboard, and clips. Joining takes one click with GitHub or a 6-digit email code. Your mic gets on-device noise removal.
- Pop-out browser windows — Right-click a browser tab to move it into its own window.
Improvements
- Cleaner settings — Every settings page is shorter and consistent: grouped navigation (App, AI, Workspace, Account), sentence-case labels, one-line descriptions or none, the same controls for the same kinds of choices, and every page translated. The AI page names each connection correctly, Secrets warns about duplicate names and values, Shortcuts follows your send key, and deleting a wallet asks first.
- Code Mode on every model — One Code Mode switch now covers Claude and Pi chats (ChatGPT, OpenAI, Muse, Grok, OpenRouter and more), and your earlier choice carries over.
- Chat titles right away — The title generates right after your first message, shimmering while it writes, and sentence-length titles are rejected.
- Orchestra defaults — The lead and captains run on Opus (Latest) at max thinking, coding subagents on Opus (Latest) at medium, explorers on Haiku (Latest).
- Captains dock — Idle captains collapse behind View all, and the captains chip only shows while they're running.
- Archived chats — Bring the archived view back from the filter menu, grouped by project.
- Per-chat full width — Toggle full width per chat from the header.
- Reveal plan file — Plan cards, the sidebar and tab menus can reveal the plan file.
- Session menu — Open in ▸ holds New Tab, New Panel, New Window and Finder; Share is gone.
- Tidier browser — Idle scratch tabs close after 10 minutes, the status bar shows the agent's active tab, and subagent tabs show in the main chat's panel.
- Lighter, steadier long chats — Idle transcripts unload from memory, image-heavy chats stay under provider limits, and "request too large" recovers instead of ending the turn. Streaming renders with fewer re-renders.
- Muse — Fewer replayed tokens per request, lock reasons for models your plan can't use, and a live model catalog.
- GitHub — Each workspace keeps its own GitHub token, the workflow scope is requested, and missing scopes are flagged.
- Messages — Better recipient search, recent conversations, report cards, and embedded email images while external trackers stay blocked.
- Home — 18 famous artist styles join the daily painting rotation.
Bug Fixes
- Cmd+N replaces the active session tab instead of opening another, and closing the last tab falls back to Home without a launcher flash.
- Markdown file links with absolute paths, including paths with spaces, open as local files.
- Commits are only attributed to the tool call that made them, and worktree commits track against the main repo.
- Pi chats fail a hung launch and report a missing working folder; chats whose folder is gone fall back to the workspace folder.
- Stop kills processes without spawning a shell.
- Subagent context in Workspace settings saves again instead of always showing Off.
- Preferences no longer overwrite a change made elsewhere while the page is open.
- Titles no longer pick up Claude Code's auto-memory from the app folder.
Your camera on stream
Features
- Camera on your live stream — Turn on your camera from the Camera section of the Live tab and viewers see you in a tile they can drag, resize (small, medium, large; double-click for big) and watch light up while you speak. Background blur and a cutout that drops your background run on your Mac, and you pick the quality from 720p up to 4K. If macOS said no to the camera earlier, Ask again brings back the real permission prompt.
- Now playing on stream — The song you're playing shows on stream with artist, play/pause state and position, and it comes back after the stream reconnects. Viewers can drag the now playing and project overlays anywhere; a throw glides into the nearest free spot and stays there.
- Stream timeline — A Timeline tab next to Live chat lays the stream out in chapters, one per prompt, with the steps under each, questions next to their answers, and a pinned Now card that catches late viewers up and jumps back to live.
- Viewers see more of your work — Your right panel, subagent and captain drill-ins, the subagent tray, and ask_user forms (one question at a time, with your picks and typed answer) now show on stream. Reactions float up and a tok/s meter rises with how fast your agents are working, counting every token your chat and its captains produce.
- Report a session — Report to Consoles... in a chat's menu sends that chat's full trace and recent logs to the Consoles team so we can fix what went wrong. Token-shaped secrets are redacted before upload.
- Code Mode for Claude Code chats — With a source connected, Claude Code chats can write one short script that calls source tools and returns only what matters, so large source results stay out of the conversation. Every source call still goes through your permission mode.
- Markdown while you type — The composer styles Markdown as you write it, and code blocks in replies render as cards with a language badge.
Improvements
- Go live is back — The Go live button sits beside Messages in the top right again, including on the dashboard and the right panel, and shows a LIVE pill while you stream.
- Live chat shows this stream — While you're live, chat shows this stream's lines under a quiet "Went live" divider, with earlier history kept above.
- Better on phones — The live viewer's controls stay up and tappable on phones, and the stream no longer goes silent on iOS.
- Subagent browser tabs — Browser tabs a subagent opens show in the panel as their own tabs, marked with who opened them, and no longer steal focus.
- Hand off in one step — Hand off sends right away as a single chip.
- Fresh home painting every day — The home dashboard paints a new scene of your city each day, in a photoreal style, with the real weather outside right now, sometimes as a photo and sometimes in a famous artist's style (credited on hover).
- Quiet timer while thinking — The thinking row shows how long the model has been quiet.
- Lighter requests on every model — Tool search now works on every Pi connection, so API-key, Copilot, Grok, OpenRouter and custom lanes stop sending every tool on every request.
- Steadier ChatGPT and Pi chats — "Server busy" errors retry instead of ending the turn, and the ChatGPT connection honors retry-after, recovers from lost replies and renews long-lived sockets.
- Subagents run on the model you name — A subagent launched on a model you asked for by name stays on it.
- Manage commands from more places — Project settings and the Processes page link to your project's commands.
- Usage counts only Consoles chats — The Usage page counts only the Claude chats Consoles ran and keeps the names of deleted sessions. Haiku 5.5 is priced, and Sonnet 5.5 cache reads are billed at the right rate.
Bug Fixes
- Sign-in never hangs — Every request to an auth server times out after 15 seconds, so a silent server can't freeze source sign-in, token refresh or the ChatGPT, Claude, Google, Microsoft and Slack logins.
- Stream redaction — Redaction no longer blurs ordinary prose, and the Apple on-device helper stops refusing and falling back to Haiku.
- Your name on the stream page — The live viewer shows your screen name instead of "New User".
Release notes in your inbox
Features
- News from Consoles in Messages — New releases arrive in your Messages inbox from the official Consoles account, marked with a verified badge. Each one is a card with the highlights, a Restart to update button once the update has downloaded, and a link to the full changelog. You can't reply to the official account, and it never wakes your agents.
- Ask for a model by name — Tell a chat to hand work to a subagent on a specific model and effort, like "send Muse 1.3 Max", and the subagent runs on it.
Improvements
- Updates arrive while the app is open — Consoles checks for updates every 4 hours and after your Mac wakes, not just at launch.
Bug Fixes
- What's new shows the real notes — After an update, What's new shows that version's actual notes instead of a generic line.
Share plans in Messages
Features
- Share plans in Messages — Type @ in a Messages reply to pick a plan from any chat in your workspace. It sends as a card with the plan's title and steps; either side can open the file or start a new chat from the plan.
Improvements
- Updates install themselves — This is the first update a signed Consoles (0.6.2) installs on its own.
Signed and notarized
Improvements
- Signed by Apple — Consoles is now signed with a Developer ID and notarized by Apple. macOS opens it without the "unidentified developer" warning, and in-app updates install by themselves from this version on. Coming from an earlier version? Download this one from consoles.ai once; after that, updates arrive on their own.
Go Live, for real
Features
- Go Live with sound — While you're live, viewers hear what you choose: your mic, a music playlist from the on-air mixer, or your Mac's audio (only Consoles' own windows), with ducking and a limiter. Songs playing in Spotify or Apple Music show up for viewers with album art and a Play on Spotify link. The On air preview shows exactly what viewers hear, and viewers start muted with Tap for sound.
- The stream matches the app — Viewers now see what you see: thinking, approval and sign-in waits, rate-limit countdowns, generated images, undo, the live model, queued follow-ups, the subagent tray, the agent's questions, plans and compaction. Streams go away when your Mac sleeps or locks and pick back up after, and viewers can load earlier history.
- Secrets stay off stream — Before anything goes out, Consoles hides API keys, tokens, passwords and your home folder path. A smart check runs on Apple's on-device model on macOS 26 (Claude Haiku elsewhere), and commands that read sensitive files like
.envare blocked. - Simpler Live controls — Pause and End sit at the top of the Live tab, with Stream, Audio and Chat in sections below. The LIVE pill in the corner opens Pause, End, Copy link and the Live tab, and going live plays a sound.
- Copy a visual as an image — Every chat visual has a copy menu: Copy as image for pasting into Slack, iMessage or Figma, or Save as PNG.
- One profile picture everywhere — Your Consoles account picture shows in Messages, on your stream, in chat and in Live now.
Improvements
- Model switches don't stop your agents — Switching models mid-chat lets the subagents and captains already running finish their work.
- Messages hover cards — Hover someone to see what they're working on, with their workspace icon, project logo and the file their agent last edited.
- Idle shows on your avatar — Your dot turns yellow after 5 minutes away, the way your contacts see it.
- A sharper Home painting — Paintings render at full retina size, and the parallax glides in one smooth layer instead of tearing apart.
- Consistent Consoles branding — App labels, bundled guides, playground examples, package descriptions, and desktop installers now identify Consoles consistently while preserving Craft integrations and upstream attribution.
Bug Fixes
- Auto-update downloads again — The update server returned "not found" for every update file, so in-app updates could never download. Fixed on the server.
- The Live tab no longer crashes after a reload while an older background process is still running.
- Live viewer page — The header and layout work for signed-in viewers.
Captains, Home, Messages, and Go Live
Features
- Captains — Turn on Orchestra and the agent you talk to can hand whole goals to captains. Each captain opens as its own chat under the main one, can work on its own branch, in its own worktree or in another repo of the project, runs its own subagents, and reports back when the goal is done. Captains can queue behind each other, settle details with their sibling captains directly, and the main agent only hears goal-level news. The Captains card above the composer and the captain rows in the sidebar show who is working, waiting on you, queued or done, with a step meter for each plan.
- Home — Consoles now opens on Home: a painting of your city in today's light, season and weather, moving in depth layers, above a task box that starts a chat in any project and cards for what's working now, pull requests with their checks, git activity, usage and recent projects. Paint a new one, travel somewhere else, or head back home from the painting. With several workspaces and nothing to restore, a workspace picker opens at launch.
- Messages — Claim a screen name and message people, projects and agents from the inbox next to the traffic lights. Chats message each other when their work overlaps; messages arrive between tool calls and read like an AIM transcript, and every chat keeps an inbox of its conversations with other chats and its subagents. You also get email at your-name@mail.consoles.ai, with a Gmail-style view, and mail from people you don't know waits until you allow it.
- Go Live — Stream the chat you're in at consoles.ai/live/your-screen-name. Viewers watch every step as it happens, chat, follow, and ask and vote on questions. Pin a note, schedule your next stream and answer questions from the Live tab in the right panel.
- Simulator — A right-panel tab that streams an iPhone or iPad simulator next to the chat (iOS only for now). With a device attached, agents on any model can boot and attach devices, install and launch the app they built, tap, swipe and type, read the accessibility tree and app logs, and take screenshots to check their work. You see every step on the same device. Opening URLs on the device follows your permission mode. Turn it off with Let agents control simulators in Settings → App → Tools; the pane still works for you.
- Subagents, rebuilt — Every backend now shares one lane system: a status line per subagent with a live token count, a tray under the composer with a report card for each lane, approvals that come up from lanes, follow-ups, resume, and messages between lanes. Subagents size themselves to the problem and launch at once, and rate-limited lanes retry by themselves or offer Retry on another account.
- Git in the composer — A git pill commits, pushes and opens pull requests without spending tokens: a drafted commit message, a commit scope that leaves out files other chats changed, live push output, Pull when a push is rejected, and PR reviews and checks with Fix, Add to PR and Sync after merge. Every turn ends with a review card listing the edited files with Undo, Review and a one-click commit. New branches name themselves, and each workspace can use its own GitHub account.
- Project commands and Processes — Save your project's commands, like dev servers and workers, in consoles.yml and run them in their own terminals. The Processes page shows everything running, and agents can start, read, restart and stop your commands.
- Image generation on every model — Agents on any backend can generate and edit images, with live partial frames while the picture paints.
- Visuals — Agents can show an HTML page inline in the chat and screenshot their own pages with the new html_preview tool.
- The full GPT lineup with an OpenAI API key — API-key connections now offer the same models as ChatGPT sign-in: GPT-6.1 Sol, GPT-6 Astra, Sol and Luna, GPT-5.6 Sol, Terra and Luna, and GPT-5.5. They get Max effort where the API supports it, the same built-in subagent roster, and Platform prices and limits (up to 922K tokens of input context).
- GPT-6.1 Sol on ChatGPT — OpenAI's new workhorse model joins the ChatGPT/Codex picker first in the list, with max effort and tool search. It has the same $2/$10 per million tokens as GPT-6 Sol with half-price cached input, and OpenAI says it nearly matches GPT-6 Astra. Built-in execution subagents now run on it, and fall back to GPT-6 Sol on accounts that don't list it yet. GPT-6 Sol stays available, labelled as the previous generation.
- Sonnet 5.5 — Claude Sonnet 5.5 joins the Claude picker, and the built-in merge agent now runs on Opus 5.5 Max (Claude) or GPT-6 Astra Max (ChatGPT).
- Jev fast mode for the browser — Turn on Jev fast mode in Settings → App → Tools and add a TypeSafe API key, and agents can hand multi-step click-and-type sequences to TypeSafe's Jev, which picks each step in about 100 ms. Values the agent types stay on your machine, and the agent takes back over whenever Jev is unsure.
- Faster, code-first browsing — Agents read pages as a compact numbered list of the controls on screen and pull data with page scripts instead of screenshots. Repeat reads of an unchanged page come from cache. Web links in the chat open as a browser tab in the right panel, and tabs can be pinned to a project or the whole workspace.
- Session tabs — Keep several chats open as tabs across the top of the window.
- A new composer — Inline code shows as a code chip while you type, pasted code is fenced, pasted Markdown becomes an editable .md badge, and double-clicking a pasted image opens it full size.
- Media and files — Videos and audio play inline with a media player, files get Save as…, Open outside and Show in Finder, and PR links show a hover card.
Improvements
- A Codex-style look — The sidebar, composer, menus and popovers use Codex's surface colors and spacing, user messages take the theme's own color, and every popover shares one surface.
- Review opens side by side — The Review button on a finished turn lands on split before/after columns. Use the diff toolbar to switch to unified; the choice is remembered. Review also reads the turn's net change from git.
- Steering like the Codex app — Queued and steering messages stack up the way Codex does, steers can carry images and files, and Stop sends a pending steer right away.
- Code Mode in a sandbox — GPT, Muse and other Pi models run Code Mode scripts in a QuickJS sandbox, now on for every Pi model.
- Accurate usage — Claude turns bill by the SDK's own cost, Pi turns bill every request, long-context API-key requests use the long-context rate, ChatGPT usage reads the right workspace, and the Usage page shows provider logos. A Use reset button spends rate-limit credits once you pass 95%.
- The window makes room — Opening the right panel grows the window instead of squeezing the chat.
- Cheaper background calls — Titles, summaries and suggestions use a cheap model at low effort, in one capped call. Titles are written once, and only change when you ask.
- Chats remember where you were — Scroll position and collapsed sections survive switching chats.
- Sources turn on when needed — On-demand sources activate as soon as an agent searches for their tools.
- Fast Mode is visible to agents — Browser context reports live Jev availability, and multi-step workflows report which steps used Jev, the small model, or cache.
- Sturdier MCP connections — Source connections reconnect, follow tool-list changes, and cancel their calls when a turn stops.
- A smaller app build — The interface bundle is much smaller, which speeds up startup and updates.
Bug Fixes
- Staying signed in — Refreshing an OAuth token in one chat no longer signs other chats out.
- No more silent Muse stalls — A dropped answer no longer leaves a Muse chat refusing every retry, and early errors now show.
- Turns finish — Runs without tools no longer hang, background turns no longer stall on a double compaction, and a crashed captain no longer blocks the reports behind it.
- Chat titles — Automatic titles fire on agentic first turns again.
- No double sends — Switching providers no longer sends a pending message twice.
- ChatGPT settings reach the model — Fast mode and verbosity now apply to ChatGPT chat turns.
- Project-local Pi config ignored — Pi sessions no longer load a .pi config from the working directory.
- File links open — Session file links and relative plan and data links open the file instead of failing to load.
- Stopped subagents are not failed subagents — Intentional stops keep their outcome and show a neutral Stopped status in the roster and detail view.
- The in-app browser works again — In v0.5.2, every model's browser calls failed with "Source "node_repl" could not be connected". Agents can drive the in-app browser again.
- Browser pages accept your clicks — Transparent browser containers no longer intercept input above the webpage.
- Browser screenshots copy to clipboard — Take a screenshot copies the image and confirms Screenshot copied instead of starting a download.
- Approving a download no longer fails the browser step — The time you spend on the approval card no longer counts against the step.
- Browser tabs keep their page — When the app window reloads, an agent's browser tab comes back on the page it was on, and switching chats no longer reloads other chats' pages.
- No more hung browser steps — Every browser command has a time limit and reports which step stopped responding, and a click on a covered element fails right away and names what's on top.
- Take control during Jev workflows — Your clicks and typing are detected while Jev waits on a model, and automation rechecks control before it types.
Breaking Changes
- No per-site browser prompts — Agents no longer ask before using a new site. Block sites in Settings → App → Browser Site Access instead.
- Built-in agents only, unless you ask — Chats launch built-in agents and @Review on their own; custom agents run when you @-mention them or add them to a blend.
- Workflow tool off — Claude Code's Workflow tool is off by default.
Browser Use for Every Model, Timeline Chat, and Live Steering
Features
- In-app browser for every model — Agents drive the browser in the side panel the way Codex does, with the same tools on Claude, GPT/Codex, Muse and every other model. A persistent, sandboxed JavaScript runtime gives the agent the page's accessibility tree, clicks and typing, Playwright-style locators, coordinate input, screenshots, dialogs, the clipboard and dev logs. Browsing runs in the background, so agent tabs keep working when the panel is closed or you switch chats. Open the panel to watch the agent's cursor at work, or ask it to show you its tab. Screenshots reach every vision-capable model and show as thumbnails in the chat, and each browser step shows the site's icon. The previous
browser_toolis still available in Settings → App → Tools. - Browser safety and control — The first time an agent acts on a new site in a session, you choose Allow for this session, Always allow, or Deny. Local dev servers never ask, and Settings → App → Browser Site Access keeps your always-allowed and blocked sites. File uploads, and downloads the agent starts, ask first unless you allow them. Click or type in a tab the agent is driving and you take over for the rest of the turn.
- Annotate pages — Turn on Annotate in the browser toolbar, click an element or drag across an area to drop numbered comment pins, then send them to the composer as text plus cropped screenshots. Hold to view the page without the pins.
- Browser mentions — Type @ in the composer to point the agent at @Browser or at one of the chat's open tabs.
- Live steering — Messages you send while an agent is working now steer the running turn on Claude, Cursor and Pi models instead of interrupting it. They reach the agent at its next tool boundary, keep their order, and anything not yet delivered goes back to the queue instead of disappearing or falsely reporting success.
- Timeline turn cards — Each turn is one timeline: a live "Working" header, then tool groups, plan steps and questions as nodes on a single line. Code Mode scripts from GPT and Muse models read as "Ran script" rows with highlighted source and readable output.
- Subagent tray — A tray above the composer lists every agent in the session with its status, model and current action. Dismiss it once no agent is working; it returns when one starts again.
- Redesigned question card — Questions from the agent show numbered options with a Recommended tag and keyboard shortcuts: 1–N to pick, Enter to submit, Esc to skip.
- New model picker and usage warnings — The model picker groups providers in one popover, context usage shows as a ring with a compaction popover, and a banner warns at 80% and 95% of a 5-hour or weekly limit.
- Streamable HTTP for MCP sources — MCP sources connect over Streamable HTTP and fall back to SSE only while connecting.
- Muse Code usage — The usage card shows the 5-hour and weekly windows of a Muse Code membership.
Improvements
- Faster chat while streaming — Finished turns no longer re-render on every streamed update, idle animations stop repainting every frame, and the Review tab no longer slows the app while a chat streams.
- Smaller debug logs — The network debug log rotates by size instead of growing to hundreds of megabytes.
Bug Fixes
- No leaked MCP connections — Switching models or providers, reverting, retrying sign-in or deleting a ChatGPT/Codex, OpenAI, custom-endpoint or Cursor session now closes its MCP connections and running tool commands instead of leaving them open until quit.
- Subagent model labels reflect the run — Agent detail headers and roster badges show the model the child actually used, not an outdated profile selection.
- Agent-name hover shows the full brief — Spawned agents show their original assigned directions on hover, including older runs whose parent task only stored the short agent name.
- Muse effort and usage — Standard Muse models keep XHigh distinct from Max effort, and the usage card keeps its last good meters when Meta briefly omits quota data.
- Chat polish — The chat header shows only the title, the Edited-files card waits until the agent is completely done, brand icons are back on the timeline, and the timeline line no longer breaks between events.
Apps, Subagents, GPT-6, and Claude Opus 5.5
Features
- Apps — Build and run custom workspace Apps inside Consoles. Apps can create focused tools, dashboards, and workflows that connect to your Sources without exposing raw credentials. Apps get a local runtime manager, a popout window, and hooks in the right panel and composer, and the agent can create and edit them from chat.
- Overview App — A built-in App gives you a high-level view of your workspace activity, sessions, and usage.
- Subagents — Delegate work to specialized agents that can run alongside your main conversation and report progress back into the session.
- Orchestrator mode — Coordinate larger tasks across agents with clearer progress and handoffs.
- iOS beta — Early iOS support is available for testing the mobile Consoles experience.
- GPT-6 family on ChatGPT/Codex — GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna join the ChatGPT/Codex model picker as “6 Astra”, “6 Sol”, and “6 Luna”, ahead of the GPT-5.6 family and GPT-5.5. All three run on the Responses Lite path with Code Mode and keep native
maxeffort instead of saturating atxhigh. Zero-config Codex subagents now use GPT-6 Sol @ max for execution and GPT-6 Luna @ medium for Explore/reading, with GPT-6 Astra @ max directing and merging. - Claude Opus 5.5 and Claude Fable 5.1 — Claude Opus 5.5 is available on direct Anthropic connections and leads the model list, with its always-on adaptive thinking handled automatically; connections that follow the catalog pick it up without changing the selected default model, and curated model lists are left untouched. Claude Fable 5.1 is selectable on Anthropic and Bedrock (us/eu/global) connections, and the Pi Anthropic lane sends Anthropic’s thinking-binding controls so mismatched replayed thinking blocks are dropped instead of failing the request. (#23)
- Claude Sonnet 5 metadata and routing — Sonnet 5 is normalized in the model picker as “Sonnet 5” with the updated speed/intelligence description, becomes the default Sonnet alias for subagents, keeps Sonnet 4.6 selectable as the previous generation, and has Bedrock US/EU/Global routing wired without changing the existing Opus 4.5–4.8 options.
- Muse (Meta) provider — Meta’s Muse Spark models (1.3, 1.2, and 1.1, including Contributor variants) are available as a first-class provider. Sign in with a Muse membership or an API key during onboarding or in Settings → AI; Muse gets its own icon and branding in the picker and chat, and runs through a native Meta Responses provider on the Pi lane.
- Goals keep their plan — Goal mode snapshots the latest to-do plan onto the goal and carries it into the next continuation, surviving turn boundaries and compaction. A new resumable “usage limited” state covers turns that end on a provider rate or billing limit — it never auto-continues and resumes from the goal chip — and the goal popover shows the latest steps and stop reason in one place. (#22)
- In-session subagents on ChatGPT/Codex — Subagent fan-out on the Pi/Codex lane now matches the Claude lane:
tasksubagents stay inside the parent session as hidden side conversations instead of appearing as separate sessions, and the session list no longer shows placeholder entries leaked by streaming races. - Codex-style right panel — The right panel now has retained workspace Files tabs, rich code/Markdown/image/PDF previews, open-outside handling for unsupported media and Office files, expandable panel width, and hidden ephemeral Side chats that inherit the parent conversation without disrupting it.
Improvements
- Safer Source access — Apps request “Use this Source” permissions and call through the Source layer instead of receiving raw API keys or OAuth credentials.
- App runtime isolation — Local Apps run in a sandboxed environment with explicit environment variables only.
- Claude Agent SDK uplifted to 0.3.280 — The bundled Claude SDK now matches Claude Code v2.1.280 behavior (up from 0.3.197 earlier in this cycle), including native Sonnet 5 awareness for aliases, context-window handling, effort defaults, first-class Claude Fable 5.1 support, and recognition of Claude Opus 5.5, which older bundles rejected as an invalid request.
- Native Codex auto-compaction and batching — ChatGPT/Codex sessions now trigger auto-compaction exactly where native Codex does — 244,800 tokens against the 258,400-token usable window — instead of 12,240 tokens early on live-discovered models, and the context meter marker matches. Responses Lite turns are guided to batch independent reads into one Code Mode
execcall withPromise.allSettled. - True usage ledger for Dashboard — Console now writes exact usage rows to each session’s
usage.jsonlfor main turns plus instrumented internal model calls likecall_llm, title generation, and summaries, and Dashboard totals are ledger-only so provider/model/connection switches are tracked without legacy header or SDK-log inference. - Review parity and reliability — Review now uses the active paired Craft code theme, supports Color or +/− diff markers, keeps hunk actions stable after context expansion, preserves staged files in Last turn, persists per-file rich-preview fallback, and rolls failed Git actions back instead of leaving optimistic UI behind.
- Live Review status — Review now refreshes from a server push whenever git status changes, with a 15-second fallback poll.
- Smoother streaming under load — Bursts of thinking and task-progress events are coalesced before they reach React, and the model picker is memoized, so long reasoning turns and large fan-outs keep the UI responsive.
- Shared command icons — Notebook command rows reuse the desktop tool-icon registry and command parser. Blender, Unreal Engine, uv, and FFmpeg join Python, npm, Bun, GitHub, and Cloudflare in the shared defaults.
Bug Fixes
- Blocked tools no longer end the turn — When Plan/safe mode blocks or denies a tool call, the model now receives the reason and continues, instead of the turn ending on a generic “No response” card that hid why.
- Background sessions no longer park forever — Three stale liveness flags could leave a session waiting on a background wake indefinitely (seen at 100+ minutes, recoverable only by restarting the app); recovery paths now check them against reality instead of trusting them.
- No phantom turns after long commands — Long-running foreground tool calls (like a slow Bash) no longer queue a background update when they finish, which used to trigger an extra continuation turn after the real turn ended.
- Idle CPU usage fixed — A re-render loop in the chat composer (for sessions without labels) kept a CPU core near 100% even with nothing running; it’s fixed, along with a Radix ScrollArea animation-frame leak.
- MCP sign-in no longer routes through upstream’s relay — Signing in to an MCP source (like Cloudflare), or to an API source that registers its OAuth client automatically, now redirects straight to the app’s local callback instead of bouncing through Craft’s
agents.craft.dorelay. The provider’s consent screen showslocalhost, and the authorization code never leaves your machine. Built-in Google, Slack, and Microsoft sign-in is unchanged. - Pi deferred session tools are discoverable again — ChatGPT/Codex sessions can now load every registered cold session control through
tool_search, including native MCP OAuth and credential prompts, instead of registering those tools in the subprocess while leaving them invisible to the model. - Pi SDK manual compaction usage reset — Manual
/compactin Pi SDK sessions now clears stale context/cache counters, carries the native compaction summary metadata through the event pipeline, and waits for the next provider-reported response to show exact post-compaction usage so the context badge no longer jumps back to the pre-compaction percentage when the turn completes or after reload. - Markdown code-fence paste handling — Long text pasted inside an explicit Markdown code fence now stays inline instead of being converted into a
pasted-textattachment, and empty code fences no longer render asundefinedin chat.
Breaking Changes
- Apple Silicon only on macOS — macOS releases are Apple Silicon only, and downloads now focus on the supported Mac and Windows builds.
Features
- Upstream Craft v0.10.3 sync from the v0.8.8 baseline, preserving Consoles-specific desktop, server, terminal, source, and branding work through a large conflict-resolution pass.
- Standalone server release line reached v1.1.0 with the same upstream baseline and headless distribution fixes.
- Expanded website release/update infrastructure for clean update routes, release manifests, R2-backed downloads, and changelog navigation.
Improvements
- Claude Fable 5 reports its full 1M context window and context usage is easier to see at a glance.
- Pi model and context-window reporting now uses resolved provider metadata instead of stale defaults.
- Large tool results are truncated more safely before proxying into model context.
- Sidebar toggle moved off Cmd/Ctrl+B to avoid colliding with common editing shortcuts.
Bug Fixes
- Per-session credential isolation prevents connection credentials from bleeding across sessions.
- Safer revert/fork behavior, including branch validation and SDK working-directory recovery across provider switches.
- Accurate context percentage after revert, compact, and provider/model switches.
- Pi image handling now defaults missing MIME types to
image/png, validates base64 more robustly, and uses the Pi SDK native image content format. - Compaction transcripts are capped to the summarization model context window to avoid overflow loops.
- PDF worker loading works reliably in dev mode.
Features
- Claude Opus 4.7 as the new default model (4.5 and 4.6 still available)
- Built-in terminal tool — agent gets a live terminal pane with notification badge and auto-open
- Provider icon on every chat bubble footer — always visible, not just multi-provider sessions
- R2 uploads migrated to S3 SDK — streaming instead of buffering
Improvements
- Claude Agent SDK upgraded to 0.2.112 (native Opus 4.7,
mcp_set_servers,WarmQueryAPI) - Pi SDK upgraded to 0.66.1 (native GLM-5.1)
- Context tooltip simplified — shows context usage + cache stats only
- Seamless message editing — recovery notifications suppressed during revert
Bug Fixes
- Cross-provider switch (Claude → Pi) no longer loses conversation history
- Tool context preserved across provider switches mid-session
- Cursor Agent validation works in dev mode
Features
- Embedded terminal with Ghosty (persistent DOM instances via node-pty)
- Docked in-app browser with ad blocking (@ghostery/adblocker)
- Right panel navigation — Files, Browser, and Terminal tabs
- GLM-5.1 model support via Z.ai provider
- Changelog page with live commit heatmap
- Release pipeline script (
bun run release)
Bug Fixes
- Resolved 95 TypeScript errors from jotai/tiptap dual-package hazards
- Session token tracking copied correctly to forked sessions
- API source credentials refresh after re-authentication
- File attachments forwarded in EditPopover mini chat
- SplashScreen unmount race and RichTextInput DOM reconciliation
- Pi model display name overrides for GLM-5.1
Features
- Deferred tool loading for Pi and Cursor providers
- Code mode (behind feature flag)
- Auto-archive workspaces after inactivity
- Configurable default directory per workspace
- Label ordering in session sidebar
- Webapp and blog with dark mode toggle
- Preserve attachments when editing a message
Upstream
- Synced Craft Agent v0.8.1 and v0.8.2
- Secure shell exec, streaming tool results, unified network interceptor
Features
- Silent auto-updates via electron-updater
Bug Fixes
- Correct function reference for onManageServers callback
Bug Fixes
- Suppress Node.js 20 deprecation warnings in CI
Features
- Unified workspace switcher: server and workspace selection merged into a single dropdown
- WebUI app and server install script from upstream
Bug Fixes
- Remove leftover sharing interfaces causing build errors
Features
- Cursor session parity: multi-turn conversations, branching, cumulative token tracking
- Universal context preservation across all providers
- Sentry source map uploads for readable production stack traces
- Auto-download Cursor Agent CLI during build
- React performance optimizations from Vercel best practices audit
Upstream
- Synced Craft Agent v0.8.0
Bug Fixes
- Preserve accumulated token usage on revert
- Include tool_use/tool_result in synthetic sessions
- Validated turn anchors for revert fork
- Resolve 30 pre-existing test failures, achieve 0-fail test suite
Features
- Google Fonts search in settings with live catalog autocomplete
- Session labels with dropdown badge in chat input
- Deep link support for sessions and automations
- Plan acceptance events through the event processor
- Context inspector redesigned with progressive disclosure
- Scroll-to-bottom floating button in chat
- Provider labels and model switch separator on chat bubbles
- Cursor Agent server integration with token counting
Upstream
- Synced Craft Agent v0.7.12
Bug Fixes
- OAuth for Figma connections
- Kanban board dialog rendering
- Inline edit bubble maintains original message size
- 1M context window opt-in activation
- Cursor and Codex icons on dark mode
- @ file mention search consistency
Features
- Kanban board view for sessions with drag-and-drop
- Cursor provider with dedicated icon and API key support
- Per-provider model picker with correct icons per section
- Searchable model lists for providers with many models
- Model capability descriptions (context window, reasoning, vision)
- Plan mode (renamed from Plan, neutral grey styling)
- Workspace name in dock menu for quick switching
- Release date sorting for models via models.dev
Upstream
- Synced Craft Agent v0.7.9 through v0.7.11
- WebSocket reconnect recovery, copilot HTTP rewrite, Docker support
Features
- Built-in media player with draggable floating window
- Compaction checkpoint cards in the chat timeline
- Frozen model labels on messages (no retroactive relabeling)
- 5-level adaptive thinking controls (Off / Low / Medium / High / Max)
- Automation conditions: time-based, state-based, and logical gates
- Amazon Bedrock provider with IAM credential support
- Connection display names and provider count badges
- Sharing links pointing to shared.consoles.ai
Features
- Mid-session provider switching with full context carry-over
- 1M context windows with reworked context usage badge
- Inline message editing with full conversation revert
- Provider icon in input bar and on chat bubble footers
- Model and connection shown on all messages via session fallback
- Message timestamp on hover of model label
Release
- Initial release of the Consoles client.