T3 Code
github.com/pingdotgg/t3code- Category
- AI Agents
- Rank
- No. 727Tools index
- Listed in
- #2 Interface with your agents
- Pricing
- Open Source
- Type
- APP
- Use case
- Agent Building · Coding
- Interfaces
- Desktop · Web · Mobile
- Builder
- pingdotgg
- GitHub
- 27.2k stars
- Latest release
- v0.0.46-nightly.20261011.2980
- Date
About
T3 Code is an open-source control plane for coding agents. It runs Claude Code, Codex, Cursor, Grok, and OpenCode through desktop, web, and mobile interfaces, keeps parallel threads visible, supports remote access, and turns reviewed agent work into commits, branches, and pull requests.
What it does
T3 Code is a client shell that sits in front of AI coding agent tools you already have installed, rather than replacing them. A background service on your machine launches and wraps those command line agents, and a desktop app, a browser tab, or a phone app all connect to that service to watch and steer the same running session. When an agent finishes editing files, the tool can turn the result into a git commit, a branch, or a pull request instead of leaving loose changes sitting on disk.
Why it's ranked here
The engineering depth here is real: the internals documentation spells out event-sourced state, transaction boundaries between intent and side effects, and specific platform traps like a shortcut-registration caching bug on Linux. But the maintainers call the project very early and warn to expect bugs, and every agent it drives needs its own separate install and login first. This is infrastructure-grade design wrapped around software its own team labels alpha stage.
What's good
It works with agent subscriptions you already pay for instead of selling a new one, and it does not lock you into one interface: the same session is reachable from a desktop app, a browser, or a phone, with install paths for Windows, macOS, Linux, iOS, and Android covered out of the box. A single shared connection layer handles reconnects and credentials once, so the three clients do not each reinvent that logic differently. Licensing is permissive, so a fork stays possible if the project changes direction.
Tradeoffs
The maintainers are explicit that this is early software with bugs, and they are not taking outside contributions beyond small fixes for now, so problems you hit may sit unresolved. Nothing here replaces the underlying agent: you still need to separately install and authenticate each provider CLI before T3 Code can control it, which multiplies setup work for anyone trying more than one. There is no hosted documentation site yet, only files in the repository.
How to use it well
This fits someone who already runs one or more coding agent CLIs locally and wants a single place to watch and direct them from a laptop, a browser, or a phone, especially if remote access matters. It is not a way to get agent access without paying for a provider separately, and given the early state and closed contribution policy, expect to work around rough edges yourself rather than getting fast upstream fixes.
Technical notes+
apps/server/src/server.ts wires together dozens of Effect Layer services in one process: provider adapters, checkpointing, source control clients for GitHub, GitLab, Bitbucket, Azure DevOps and Forgejo, auth, and telemetry, all composed with Layer.provideMerge chains. apps/desktop/src/main.ts does the equivalent for the Electron shell, layering window, backend pool, SSH tunneling and update-check services around the same server. The connection, relay, rpc and authorization index modules under packages/client-runtime/src re-export the session, retry and credential logic that docs/internals/connection-runtime.md describes as shared by web, desktop and mobile so reconnect behavior does not diverge per client. packages/contracts/src/index.ts is the single schema module every surface imports for the wire protocol. docs/internals/overview.md documents an event-sourced orchestration engine where commands commit as events before any side effect runs, and checkpoints stored as hidden git refs rather than commits. docs/user/permission-modes.md lists four approval modes (Supervised, Auto-accept edits, Auto, Full access) with per-provider caveats, since Auto only has automatic review on some providers. CONTRIBUTING.md states PRs are auto-labeled with a vouch trust tag and unvouched external PRs are expected by default. LICENSE is MIT, and package.json shows a pnpm-managed TypeScript monorepo pinned to Node ^24.13.1.
Observed
- License
- MIT (LICENSE, T3 Tools Inc.)
- Language and packaging
- TypeScript monorepo managed with pnpm workspaces, declared package manager pnpm@11.10.0
- Runtime requirement
- package.json engines field requires Node ^24.13.1
- Interfaces
- command line install and server, Electron-based desktop app, web app, and native mobile apps for iOS and Android
- Install surface
- shell and PowerShell install scripts, one-off run via npx, winget, Homebrew cask, and an AUR package
- Platform support
- Windows, macOS, and Linux for desktop; iOS and Android for mobile
- Providers driven
- Codex, Claude Code, Cursor, Grok Build, OpenCode, and Antigravity, each needing its own separate CLI install and login
- Source control integrations
- adapters for GitHub, GitLab, Bitbucket, Azure DevOps, and Forgejo present in the server source
- Contribution policy
- repository states it is not actively accepting external contributions beyond small, focused fixes
Read from README.md, package.json, AGENTS.md, apps/server/src/server.ts, apps/desktop/src/main.ts, packages/client-runtime/src/connection/index.ts, packages/client-runtime/src/relay/index.ts, packages/client-runtime/src/rpc/index.ts, packages/client-runtime/src/authorization/index.ts, packages/contracts/src/index.ts, docs/internals/connection-runtime.md, docs/internals/overview.md, docs/user/permission-modes.md, CONTRIBUTING.md, LICENSE.
What it can do
Orchestrate multiple coding-agent CLIs (Claude Code, Codex, Cursor, Grok Build, OpenCode) from a single interface
Coding-agent CLI commands → Agent responses/actions
Intel on T3 Code
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.