T3 Code: orchestrating agents from several providers in one place
T3 Code is a free, open-source interface to control Claude Code, Codex, Cursor and other agents from one place. What it does and when it makes sense to use it.
What is T3 Code, and what is it not?
T3 Code is not another agent or a model. It is the control layer that sits on top of agents that already exist. A server runs on your machine and the clients (an Electron desktop app, a web app and mobile apps) connect to it. The agents remain each provider’s own: T3 Code uses their command-line tools and your accounts; it does not replace them.
For a designer who works with code, this matters for a practical reason: you can launch the same task with two different agents and compare the results in one place. It is a concrete way to see what agent orchestration means: the layer that decides which agent does what.
Which providers and platforms does it support?
According to the repository README, T3 Code works with these agents:
- Codex (with
codex login). - Claude Code (with
claude auth login). More in Claude Code. - Cursor (with
agent login). More in Cursor. - Grok Build (with
grok login). - OpenCode (with
opencode auth login). - Google Antigravity: enabled in the app settings with “Sign in with Google”, with no CLI to install.
You must install and sign in to at least one before you start. The official site says more agents ship frequently, so this list can change. On the platform side, there is a desktop app (installable from GitHub Releases, winget, scoop, Homebrew, a .deb package, AUR and Nix), a command-line server, a web app at app.t3.codes, and iOS and Android apps that work as a remote control.
What can you do with it?
The official documentation describes, among others, these features:
- Threads with permission control. Each thread has its own mode: Supervised (asks for approval before running commands or editing files), Auto-accept edits, Auto and Full access. Watch out: Full access, which runs and edits without asking, is the initial default. Switch to Supervised if you do not want an agent touching files without telling you.
- Version control. A thread can work in its own worktree and branch. From the app you can review the diff, generate commit messages, push and open pull requests. It supports GitHub, GitLab, Bitbucket, Azure DevOps, Forgejo and Gitea.
- Remote access. From a phone, a browser or another computer, through T3 Connect, a local network, Tailscale or SSH.
- Outside agents over MCP. Other agents (Claude Code, Codex, ChatGPT) can connect to the server over MCP. The default permission is read-only.
- Multiple accounts. You can use several Codex and Claude accounts.
How is it installed?
According to the README, there are three paths:
- Command line. On macOS and Linux, an install script from t3.codes; on Windows, the PowerShell equivalent. Then run
t3. - Without installing. With
npx t3@latest. - Desktop app. From GitHub Releases or a package manager.
Before any of them, install and sign in to an agent. Without one, T3 Code has nothing to control.
How much does it cost, and what are its limits?
| Aspect | What it means |
|---|---|
| License | MIT, open source |
| Own cost | Free. It does not resell tokens: you pay each provider what you already pay |
| Where it runs | On your machine. If it turns off or loses the network, remote work stops |
| Maturity | The README warns it is a very early project with bugs. When this article was reviewed, the latest version on GitHub Releases was 0.1.6 |
| Remote access | T3 Connect allows up to 3 machines per account; a paid T3 Chat plan raises the limit |
| Contributions | The maintainers are mostly not accepting outside contributions for now |
The security point that weighs the most is remote access. The documentation asks you to treat pairing links like passwords, and warns that, by default, a paired device receives all permissions, including the terminal.
When should you NOT use it?
- When you need work to continue with your computer off. That requires cloud agents or background agents hosted outside your machine.
- When your team needs a stable tool for deliveries with fixed dates: the project itself says it is early.
- When you use only one agent. A control layer adds more with several.
- When you cannot read a diff. T3 Code makes review easier, but it does not do it for you.
Example
Illustrative example: a designer wants to know which agent respects her design system better. She has Claude Code and Codex signed in on her laptop and opens T3 Code with her prototype’s repository. She creates two threads, each with its own worktree, in Supervised mode, and asks for the same thing in both: “Create the empty state of the projects list with the EmptyState component and the tokens in tokens.json”.
She reviews both diffs in the app. One created new styles and the other reused the existing component. She keeps the second, commits from the thread and discards the other. One test does not let her generalize, but the side-by-side comparison gave her something concrete to show the team.
Common mistakes
- Installing T3 Code without a signed-in agent.
- Leaving the permission on Full access without realizing it is the default.
- Launching several agents on the same files without separating worktrees.
- Leaving long tasks on a computer that is about to turn off.
- Sharing a pairing link as if it were an ordinary URL.
- Accepting commits without reading each thread’s diff.