This is a fork of T3 Code that integrates a few additional features, including:
- Whiteboard: A tldraw whiteboard that sits side-by-side with T3 Code, allowing users to interact with agents (or other users) by editing and pointing to the whiteboard. Both the user and the agents can conveniently write links that point to a rectangular region on the whiteboard. Agents can also draw Mermaid diagrams on it, such as flowcharts and statecharts.
- When the whiteboard's config maps T3 Code's hostname to
"app", that hostname opens T3 Code itself, and the project's board opens as a Whiteboard tab in its right panel (+ → Whiteboard). There, Settings → Right panel lets each browser hide items of the + menu or change their letters, over machine defaults set in the whiteboard's config.
- When the whiteboard's config maps T3 Code's hostname to
- Better support for manually switching accounts mid-session:
- After hitting the usage limit of an account, the thread now retries after 2 minutes and 30 minutes in addition to the default retry behavior of T3 Code. This allows the thread to recover earlier if the user manually switches the account.
- New Codex/Antigravity logins are picked up mid-session without a full restart.
- Display mode per thread: Right-click a thread and set Display mode to "Chat mode" to hide the agent's work and keep the chat simple: you see just your messages and the replies the agent sends to you with its
send_to_usertool, and the agent answers each message where you wrote it, so a question in the main list gets its answer in the main list. "Normal mode" shows everything again. The mode applies in the web and desktop apps, not in the mobile app.- To follow up on one of the agent's messages without filling the main list, point at it and choose Reply. The replies show in a Replies tab in the right panel (+ → Replies, or the reply line under the message), which has its own reply box, and the agent answers you there. Under a message with replies you see how many it has, how many are new, and a few generated words on what they are about. The agent names the message it answers by its id, with
send_to_user. A Cite in the main list on a message that already has replies opens those replies and puts the quote in their reply box, unless the main box or the open reply box already has text; then the quote goes there, the main box first. A Cite on any other message goes into the main box. An agent reads the thread as you see it with itst3_thread_user_viewtool, and to brief another agent in your own words it passes on the result's last line, which shows that agent the same messages and says if they changed since. - In Chat mode the agent sends one reply per message you type, and before a long task it may first send a short "on it" note; a thin line parts two agent messages in a row. To say more, it rewrites its reply in place with its
replace_send_to_usertool. A rewritten reply shows a small "Edited" mark, and in a sub-thread it counts as new. - "Enable Dashboard", beside Display mode in the same menu, gives the thread a Dashboard in either mode: a Dashboard tab in the right panel (+ → Dashboard) with six status slots that the agent writes with its
dashboard_settool and reads withdashboard_get. Each slot shows how long ago it changed, and new text stays highlighted until you click the slot. Select text in a slot to quote it with Cite, as you would in a message. - Settings → Threads → New threads → Display mode picks the mode for threads you start from the message box in the web or desktop app. Threads that agents, scheduled tasks, side chats or the mobile app start open in Normal mode, and existing threads keep their mode.
- To follow up on one of the agent's messages without filling the main list, point at it and choose Reply. The replies show in a Replies tab in the right panel (+ → Replies, or the reply line under the message), which has its own reply box, and the agent answers you there. Under a message with replies you see how many it has, how many are new, and a few generated words on what they are about. The agent names the message it answers by its id, with
- Split the right panel in two: drag a right-panel tab onto the left or right part of the panel's content and drop it to get two panes side by side, each with its own tabs, as with Windows Snap. Drag tabs between the panes, drag the line between them to resize (double-click resets it), and close or move away a pane's last tab to go back to one pane. The tab menu has Split right, Split left, Move to the other pane and Unsplit. New tabs open in the pane you last used. A narrow window shows all tabs in one strip and keeps the split for when there is room. A press anywhere on a tab now switches to it at once.
- Copy buttons on pages in the preview browser now put the copied text into your own clipboard.
- Two preview browser modes: Settings → Browser → Preview browser picks the browser the server runs for preview tabs. T3 native (fast) is the default. It runs chrome-headless-shell, Chrome's small build that runs without a window. Chrome (slow) runs full Chrome, for pages that don't work in T3 native. On platforms where T3 has no Chrome for Testing build (Google's Chrome build for test automation), both choices run T3 native. Switching closes the open preview tabs.
- Compact when idle for Claude threads: a choice in Settings → Providers → Advanced → Compact idle threads. Besides compacting before the prompt cache expires, it compacts a large idle Claude thread (default 400K tokens or more) a set time (default 15 minutes) after the last message or answer from a user or a bot, so the compaction runs while no one waits for a reply. The token counts and the wait are set right under the choice.
- Prune idle worker threads:
t3 prune-worker-threadsdeletes the messages, turn items, nodes, plans and runtime requests of idle worker threads. A worker thread is idle when it has no open run and, in the last 7 days, was not created, had no run requested or ended, and had no turn item or message updated. A worker thread is a subagent thread that an agent started. Subagent threads that a bot starts are workers too. Threads a person opens, bot threads themselves and a person's forks are never workers. An idle worker keeps its rows while recovery still needs the worker, or while a thread that keeps its own rows names the worker as parent or fork source. Pruned threads keep their thread row, runs, parent link and pull request links, so they still show up and open with no messages. The command runs only when you call it, so run it daily to keep worker transcripts from piling up. Other threads still grow the database. With--dry-run, the command only prints how many threads it would empty and how many rows it would delete. You can run the command while the server runs. - One way to start subagents in Claude and Codex threads: T3 Code starts Claude and Codex sessions without the provider's own subagent tools: Claude's
Agent,Workflow,SendMessageandListAgents, and Codex'sspawn_agentwith its other collaboration tools. The agent starts every subagent, same provider and model included, with T3 Code'sdelegate_tasktool, so each child is a T3 Code thread that you can open and follow. - Files remember where you stopped reading: a file in the right panel opens where you left it when you switch back to its tab, close and reopen it, or reload the app. This covers code, Markdown, PDFs, large text files, CSV and TSV tables, HTML pages, Word and PowerPoint files, the first sheet of an Excel or ODS spreadsheet, notebooks, Mermaid diagrams and chat attachments; video and audio come back paused at the time you left. If the file changed, it opens at the same place, or at its start or end when that place is gone. Places stay in this browser for 30 days. A link to a line still jumps to that line once. PDFs in the right panel open in T3 Code's own PDF viewer, also in Chrome and Edge on the desktop. Image zoom, fonts, archives, SQLite files, ODT, ODP, RTF and EPUB files, HTML attachments and the other sheets of a spreadsheet still open at the top. A place in a notebook cell after the first 20 opens at the end of cell 20.
- Reflink worktrees on Linux: Settings → Worktrees → Reflink worktrees (on by default) clones new worktrees copy-on-write from a cache of recent checkouts on the worktree location's drive, so worktrees share disk space until files change. Drives that cannot clone, such as ext4, use Git's normal checkout.
- FYI messages between agents: an agent can send another thread an FYI, a note of at most 16,000 characters that needs no action and starts no turn. The receiving agent gets it with its next turn; a running Claude Code thread gets it at its next tool call. An FYI counts as delivered only when the turn that carries it finishes: a turn that fails, is stopped, restarts or is reverted gives it back, so the agent may see it twice. When a thread has an FYI waiting and has been idle for 6 hours, T3 Code wakes it to deliver its FYIs, unless the thread is settled, archived, waiting for your answer or holds queued messages; a snoozed thread waits until its snooze ends. The 6 hours count from the thread's last turn, or from the reset of a usage limit that stopped it on the provider it still uses (or from the end of a snooze that ends after that reset), so an FYI to a thread that has been idle longer wakes it within minutes. A thread whose usage limit has no known reset time, or whose provider fails every turn, gets a failed wake every 6 hours while an FYI waits. A wake carries up to 16,000 characters of FYIs. The rest, and any that a wake gives back because it failed or you stopped it, get another wake after 6 more idle hours. A wake is a new turn, so Resume no longer continues a turn that ended before it. Settle or archive a thread to stop its wakes. The timeline marks these messages "FYI · waiting", "FYI · sent" while the turn that carries them runs, and "FYI" once it finished.
- Open server files by their path: in a browser signed in to T3 Code at its own address, open
<T3 address>/<absolute path>(such ashttps://t3.example.com/home/ana/work/report.html) to see that file read-only, for files under the server user's home folder and temp folder. A session needs thefilesystem:readpermission, and a signed-out browser gets a sign-in page. Text and code of any type show as plain text, and images, PDFs, audio and video show as themselves. A folder shows itsindex.html, or else a list of its files. Links ending in:lineor:line:colopen the file. Names that start with., and T3 Code's own data folder, are not served. An HTML page runs in a sandboxed frame at the same address: its scripts run but cannot act as you in T3, and it loads files from its own folder and below for an hour, though its scripts cannot fetch them. A page directly in the home or temp folder loads no other files. An HTML page linked from another site, or any HTML page over plain HTTP, first shows an Open link. - HTML pages from disk block their own network calls: an HTML file that the file viewer, "Open in browser" or a server file path shows from disk runs its own scripts and loads scripts, styles, fonts and images over HTTPS. It cannot use fetch, XHR, WebSocket, EventSource, beacons or form posts, its frames cannot load other addresses, and audio and video play only from its own files or inline data. A page that fetches its data while it runs needs that data inline. An allowed load can still carry data in its URL, so these limits do not make an untrusted page safe.
- Enabling math rendering by default.
- Other small fixes.
The following is the README from upstream.
The open-source control plane for coding agents.
Download · iOS · Android · Web · Docs · Discord
T3 Code runs the coding agents on your machine and gives them one fast interface. Start a thread at your desk, then approve the next step from your phone.
It works with Claude Code, Codex, Cursor, OpenCode, Grok Build, Antigravity, Kiro, Pi, Muse Code, fx, and any ACP agent. If an agent is set up on your computer, T3 Code can control it.
- Your subscriptions. T3 Code does not resell tokens. It drives the agents you are already signed in to.
- Every agent in one place. Change models in the middle of a thread. Let Claude hand a task to Codex.
- Desktop, web, and phone. One server, and every client shows the same threads. There are apps for iOS and Android.
- Remote ready. Run agents on any machine. Connect over your local network, Tailscale, or T3 Connect.
- Git built in. Each thread can work in its own worktree. Review the diff, then commit, push, and open a pull request with one button.
- Fast. Performance is a feature here. We measure it, and we treat a dropped frame as a bug.
- Open. MIT licensed. If you don't like something, fork it.
Download T3 Code for macOS, Windows, or Linux. You can also use a package manager:
| Platform | Command |
|---|---|
| macOS | brew install --cask t3-code |
| Windows | winget install T3Tools.T3Code or scoop install extras/t3code |
| Arch Linux | yay -S t3code-bin |
| Nix | nix profile add github:NixOS/nixpkgs/nixos-unstable#t3code |
| Debian, Ubuntu | .deb from Releases |
This installs the t3 command. Run t3 to start the server and open the web app.
curl -fsSL https://t3.codes/install.sh | shOn Windows, in PowerShell:
irm https://t3.codes/install.ps1 | iexTo try it once without an install, run npx t3@latest.
Get the app from the App Store or Google Play. Then follow remote access to connect it to your computer.
T3 Code needs one agent that is installed and signed in. For example, install Claude Code and run claude auth login, or install Codex CLI and run codex login.
Provider setup has the steps for each agent. Install T3 Code covers nightly builds, AppImage, WSL, and NixOS.
Nothing. We built T3 Code because we wanted the best possible development experience with agents. We were inspired by existing solutions like the Codex desktop app, Conductor, Claude Desktop and Cursor Glass, but none met our bar.
We wanted something performant, remote-ready, and truly open. If we ever go the wrong direction, we want you to have everything you need to fork and build the tool that you want.
- Install and first run
- Remote access from a phone or another machine
- Permission modes
- Source control
- Connect outside agents over MCP
- FAQ
All docs has the full list.
We have limited review capacity. Small, focused fixes are the most likely to be merged. Features need an approved discussion first.
- Read CONTRIBUTING.md before you report a bug or open a pull request.
- Put feature requests in Ideas.
- Build from source with the development runbook. The architecture overview explains how it works.
- Get help in Discord.