T3 Code Explained
A Free GUI for Codex, Claude Code and Cursor Agents (2026)
T3 Code is a free MIT GUI to run Codex, Claude Code and Cursor from one screen. Install commands, phone remote access, and a comparison with Herdr, Conductor.
T3 Code is a free, open-source GUI for operating the coding agents that run on your own machine (Codex, Claude Code, Cursor, Grok Build, OpenCode, Google Antigravity and more) from a single screen. The official README calls it an "agent harness control surface," and it runs as a desktop app (Electron), a web app, and iOS and Android apps. It is MIT-licensed, and T3 Code itself costs nothing (you still pay for each agent's own subscription).
It is built by pingdotgg (Theo Browne and team), has about 53,000 GitHub stars, and ranked near the top of GitHub trending in November 2026. This article uses the GitHub README, the official docs (docs/user) and the Releases page as primary sources to cover what it does, how to install it, the shortest way to start, how phone-to-desktop control works and its caveats, and how it differs from similar tools (as of November 2026).
What T3 Code is — the key points in 3 lines
- What it is: a control screen for coding agents on your machine or server. T3 Code is not an agent engine (harness) itself; it uses the Claude Code and Codex harnesses
- Where it runs: desktop app, web app (app.t3.codes), iOS and Android apps, and the t3 command
- Cost: T3 Code is free and MIT-licensed. You still pay for the agents (Claude, ChatGPT/Codex and so on) as before
The official FAQ frames it as: "Chef = LLM, Kitchen = harness, T3 Code = the restaurant that lets you bring whatever kitchen you want." In other words, T3 Code is not a new model or a new agent. It is a control layer and remote connection on top of agents you already use. Instead of running claude or codex directly in a terminal, you drive the same agents from a GUI, side by side, and from your phone.
What it can do
| Feature | What it does (per the official docs) |
|---|---|
| One place for many agents | Switch providers and start threads (units of work) in the same screen. You can send the same prompt to several models at once (Shift-click; requires a Git project) |
| Threads and worktrees | Start a new thread per task, and choose New worktree when changes need a separate branch and directory |
| Permission modes | Four modes: Supervised, Auto-accept edits, Auto, Full access. The initial default is Full access |
| Diff review and Git | Review each turn's changed files and diffs, or review commit by commit. Works with GitHub, GitLab, Bitbucket, Azure DevOps and other hosts, including creating pull requests |
| Remote access | t3 up lets you use your machine from other devices (T3 Connect, LAN, Tailscale, SSH) |
| Browser and devices | You and the agents share the same browser tabs. An iOS Simulator or Android Emulator can be shown next to a thread so you can watch an agent verify mobile work |
| MCP | The README links to instructions for connecting Claude Code, Codex, ChatGPT and other agents over MCP |
The pace of change is fast, and the README states up front that the project is "very early" and to "expect bugs." It also says large outside contributions are not being accepted. For real work, keep that in mind and always review the diff yourself.
Supported agents — per the README and official docs
The README lists seven providers, and the install doc (install.md) adds Kiro, Pi and Muse Code. The table separates the two. In every case the agent must already be installed and logged in on the machine where T3 Code runs (your PC or a server).
| Provider | Setup (as documented) | In README |
|---|---|---|
| Codex | Install Codex CLI and run codex login (connecting with ChatGPT is also possible) | Yes |
| Claude (Claude Code) | Install Claude Code and run claude auth login | Yes |
| Cursor | Install Cursor CLI and run agent login | Yes |
| Grok Build | Install Grok Build CLI and run grok login | Yes |
| OpenCode | Install OpenCode and run opencode auth login | Yes |
| Google Antigravity | No CLI needed. Enable it in Settings, then Install Antigravity and Sign in with Google | Yes |
| fx | Install fx, then run fx login on the machine running the T3 Code server | Yes |
| Kiro | Install Kiro CLI and run kiro-cli login | Docs only |
| Pi | Install Pi (Node.js 22.19 or newer), run pi, then /login | Docs only |
| Muse Code | Install Muse Code on the server and run muse login | Docs only |
Enable providers under Settings → Providers after choosing the environment (machine). The provider CLI must be on the server's PATH; if T3 Code cannot find it, set the Binary path in the provider settings. Pi is the minimal harness covered in our Pi (pi.dev) guide, and T3 Code can select it.
According to the FAQ, T3 Code operates within the terms of Claude Code, Codex and Google Antigravity, so you can use your existing subscriptions. The FAQ also notes a known issue: because Claude is driven through Anthropic's Agent SDK, the 5-hour usage limit drains faster than with Claude Code directly (reported to Anthropic as a bug on its side). Check this if you watch your Claude limits closely.
Install and pricing
There are three ways to get it: the t3 command, the desktop app, and the mobile app. The commands below come from the README and docs/user/install.md.
# macOS / Linux (command-line version)
curl -fsSL https://t3.codes/install.sh | sh
# Windows (PowerShell)
irm https://t3.codes/install.ps1 | iex
# Try it once without installing (needs Node.js)
npx t3@latest| OS / form | Install (per the README) |
|---|---|
| Windows (desktop) | winget install T3Tools.T3Code or scoop install extras/t3code |
| macOS (desktop) | brew install --cask t3-code |
| Debian / Ubuntu | Download the .deb from GitHub Releases, then sudo apt install ./T3-Code-*.deb |
| Arch Linux | yay -S t3code-bin (stable) / yay -S t3code-nightly-bin (nightly) |
| Nix | nix profile add github:NixOS/nixpkgs/nixos-unstable#t3code |
| iOS / Android | The "T3 Code" app on the App Store / Google Play |
The command-line installer places t3 in ~/.local/bin. There is no t3 executable for Intel Macs (the desktop app is available); the docs say to build from source to run a server there. The desktop app bundles the t3 command, which you can install from Settings → About.
The FAQ answers the pricing question with "Nothing." There is no paid T3 Code plan; only the agent costs apply. T3 Connect, the remote connection service, is also free for now, and lets each account link up to 3 machines (phones and browsers do not count). The FAQ adds that the team may change course if the bill becomes unacceptable, so watch for future changes. A paid T3 Chat account raises the limit to 10 machines (Pro) or 100 (Premier).
Shortest path to using it — first launch to first thread
- 1. Start it: with the CLI, run t3, which starts the server and opens the local web app. With the desktop app, just launch it
- 2. Enable an agent: in Settings → Providers, enable an agent that is installed and logged in (for example, Claude after claude auth login). Installation and login happen on the machine that runs the agent
- 3. Open a project: open the repository you want to work in. Threads can also be created without a project, in which case a dedicated folder is made under ~/.t3/scratch
- 4. Create a thread and give instructions: New thread starts a task. Choose New worktree to isolate code changes. Cmd+Enter (Ctrl+Enter on Windows and Linux) sends and immediately opens another draft so you can work in parallel
- 5. Pick a permission mode: the default is Full access (runs without approval), so until you are comfortable, switch to Supervised, which asks for approval on commands and file changes
- 6. Review the diff: check each turn's changes and diff, then commit or open a pull request
To keep it running in the background on macOS or Linux, use t3 service install. Update with t3 update, and see all commands with t3 --help. The default of Full access is the setting most likely to cause accidents, so review it first.
Controlling your desktop agents from a phone — how it works and cautions
The mobile app is a remote control. The real work (reading and writing code, running commands) happens on your PC or server. The official FAQ says the same: the phone is a remote, and the computer doing the work must stay on. The full procedure is in docs/user/remote-access.md.
| Method | How it works | Good for |
|---|---|---|
| T3 Connect | Connects from other signed-in devices without router port forwarding. Run t3 up on the host and choose T3 Connect | Easy access while away |
| LAN / pairing | t3 up --lan and similar; connect devices on the same network with a QR code or pairing URL | Home or office network |
| Tailscale | t3 serve --tailscale-serve or t3 pair --tailscale for HTTPS over Tailscale | If you already use Tailscale |
| SSH | Add an SSH host as an environment from the desktop app. Projects and credentials stay on that host | Running on a dev server |
The official docs spell out several security points. First, in LAN mode T3 Code accepts connections from your local network and tailnet, so the pairing code is all that keeps other devices out. Second, create a fresh one-time pairing link for each device. Third, agent logins, credentials and work stay on the host machine, not the phone. Fourth, the hosted web client app.t3.codes connects directly to your server and needs an HTTPS endpoint.
In practice, the combination with permission modes matters. Leaving a remotely reachable host on Full access means a lost device or leaked connection data translates directly into command-execution rights. For hosts you reach from outside, use Supervised or Auto-accept edits, and revoke sessions for devices you no longer use in Settings → Sharing & access (or t3 auth session revoke <id> on the CLI).
Recent updates — November 2026 releases
GitHub Releases carries both stable builds and nightlies (preview builds that appear almost daily). As of November 9, 2026, the latest stable release is v0.1.7 (November 4), preceded by v0.1.6 (October 31). Nightlies in the v0.1.8 line were still shipping on November 9, several times a day. As the README says, it is at an early 0.1 stage.
| Version | Date | Main changes (per Releases) |
|---|---|---|
| v0.1.8-nightly.20261109.4399 | November 9, 2026 | Suggests a follow-up prompt after Claude and Codex turns; simpler CLI setup where npx t3 can install the t3 command; agents can request sudo passwords privately; and more |
| v0.1.8-nightly.20261109.4392 | November 9, 2026 | Free unused database space from Settings; faster worktree thread creation on macOS (beta); drag search results into the composer; and more |
| v0.1.8-nightly.20261108.4353 | November 8, 2026 | Antigravity CLI usable as a custom executable |
| v0.1.7 (stable) | November 4, 2026 | Stable release |
Nightly contents change daily, so use the stable build for real work and choose nightly only to try new features. The docs say to set the environment variable T3CODE_CHANNEL=nightly at install time to get the nightly train.
How it differs from existing tools — Herdr, Conductor and the first-party apps
Running several coding agents in parallel has both terminal-style and GUI-style options. The table covers only what could be confirmed on each tool's official site or in the T3 Code FAQ; "not stated" means the primary source did not say.
| Tool | Form | Supported agents | Price / OSS | Notes |
|---|---|---|---|---|
| T3 Code | GUI (desktop, web, iOS, Android) plus t3 command | Codex, Claude Code, Cursor, Grok Build, OpenCode, Antigravity, fx (plus Kiro, Pi, Muse Code in the docs) | Free, MIT (agents paid separately) | Mobile and remote access are central |
| Herdr | Native app plus terminal (TUI) | Any agent that runs in a terminal (status display needs OSC 7501 support) | Free, Apache 2.0 (official site) | macOS (Apple silicon), Linux, Windows; other machines over SSH |
| Conductor | GUI (Mac) | Claude Code, Codex, Cursor (official site) | Not stated on the site | Runs agents in parallel in separate workspaces on a Mac |
| OpenAI Codex app | OpenAI's first-party app | Codex (its own) | Not stated here | Named as an inspiration in the T3 Code FAQ |
| Claude Code desktop / web | Anthropic's first-party app (Code tab) | Claude Code (its own) | A Claude plan is required | The T3 Code FAQ lists Claude Desktop as an inspiration |
| Cursor | Editor / CLI | Cursor (its own); Cursor CLI can be connected as a T3 Code provider | Paid plans exist | Cursor agents can be operated from T3 Code |
In one line, T3 Code leans toward an agent-agnostic control screen with remote access from your phone. If you want to organize parallel runs around the terminal, multiplexers such as Herdr or Corgi are closer, and if you only want Git worktree management, worktrunk is a lightweight option. First-party apps are tuned for their own agent but generally do not handle other vendors' agents.
Who it is for, and who it is not
- Good fit: people who use several agents such as Claude Code and Codex / people who want to check progress and send instructions from a phone to agents on a home PC or dev server / people who like reviewing diffs in a GUI / people who want an open-source tool they can fork
- Not a fit: people who want to stay entirely in the terminal (Herdr and similar are closer) / teams that put stability first and do not want early-stage 0.1 software / people who use a single agent and are happy with its official app
- Check before adopting: the default permission mode is Full access; manage a remotely reachable host (revoke pairings); and the faster 5-hour limit drain when Claude runs through the Agent SDK
Sources
- T3 Code (GitHub README)
- Releases
- Install (docs/user/install.md)
- Remote access (docs/user/remote-access.md)
- Permission modes (docs/user/permission-modes.md)
- Threads (docs/user/thread-sidebar.md)
- FAQ (docs/user/faq.md)
- t3.codes / app.t3.codes
- Herdr official site / Conductor official site
Is T3 Code free?
Yes. The official FAQ answers "Nothing," and T3 Code itself is free and MIT-licensed open source. You still pay for the agent subscriptions or API usage (Claude, Codex and so on). The T3 Connect remote service is also free for now, limited to 3 machines per account.
Can I use my existing Claude Code or Codex subscription?
Yes. The FAQ says you can use your Claude Code subscription as is. Install and log in to the agent first, and T3 Code then drives it. The FAQ also notes a known issue where Claude, driven through the Agent SDK, drains the 5-hour limit faster.
Does T3 Code replace Claude Code?
No, it sits on top. The FAQ explains that T3 Code is not a harness; it uses the harnesses of Claude Code, Codex and others and does not work without one. You keep Claude Code installed and use T3 Code as the control screen.
What do I need to use it from my phone?
On the PC or server running the agents, run t3 up and allow access through T3 Connect (or LAN or Tailscale). Install the T3 Code iOS or Android app and connect. The work happens on your machine, so it must stay powered on and reachable.
Is remote access safe?
The official docs state that in LAN mode the pairing code is what keeps other devices out, and that pairing links should be one-time per device. They do not make a general safety guarantee, so for a host reachable from outside, use the Supervised permission mode and revoke sessions for devices you no longer use.
Should I use the stable build or nightly?
Stable is the safer choice for work. As of November 9, 2026 the latest stable is v0.1.7 (November 4), while the v0.1.8 nightlies update several times a day. The README itself says the project is very early and may have bugs. Pick nightly only to try new features early.
Related free tools (no sign-up, instant results)
Feel free to contact us
Contact Us