Drive your T3 Code threads from ChatGPT: add its MCP server as an app, and keep it Read only until you need Supervised

T3 Code, the open-source app that runs Codex, Claude Code and other coding agents side by side, now lets agents it did not start connect to its MCP server. Julius (@jullerino) from the T3 team said on Oct 7 that it works with any connector, so you can add it to ChatGPT, Claude or Grok and drive your threads from there. We checked the setup against T3 Code's outside agents guide.

#What changed

Each T3 Code environment exposes an MCP endpoint at its address followed by /mcp. An outside agent signs in once with OAuth, and you choose what it may do. With Read only access it can read projects and threads and see which providers and models are available. With Supervised or Full access it can also start, message and stop threads in every project, and a thread it starts cannot run with more permissions than the mode you picked.

#Who is affected

This is for people who already run Codex or Claude Code inside T3 Code and would rather talk a task through in ChatGPT before handing off the coding. It also covers Codex CLI and Claude Code in your own terminal. A hosted agent like ChatGPT has to reach your environment from the internet, so it needs the T3 Connect address; a Tailscale address only works from inside your own tailnet.

#What to do now

1. In T3 Code, open Settings → Connections, open the saved environment's menu, and choose Copy MCP URL. For ChatGPT, make sure it is the T3 Connect address.
2. In ChatGPT's apps settings, create a new app with a custom MCP server, paste the URL, and choose OAuth. You do not need a client ID or secret.
3. Approve the sign-in page only if you just started it. Enter a pairing code from Settings → Connections, or run t3 auth pairing create on the host.
4. Leave access on Read only at first. Switch to Supervised only when you actually want ChatGPT to start or steer threads.
5. Try a simple handoff, suggested by @jayzhoupro: discuss a bug in ChatGPT, ask it to open a T3 thread with the brief you agreed on, then check progress from the same chat while the code work stays organized in T3.
6. From the Codex CLI the same server works with codex mcp add t3 --url https://<environment-address>/mcp followed by codex mcp login t3.
7. Revoke an agent any time under Settings → Connections. A sign-in lasts 30 days before it asks you to approve again.

#What is not confirmed

• The guide lists reading, starting, messaging and stopping threads. It does not say whether you can answer a thread's permission prompts over MCP; a user asked Julius and we found no answer yet.
• When you approve a hosted agent, T3's guide warns that anyone who runs that service gets the access, so think before granting Supervised or Full access to a cloud client.
• We did not run the ChatGPT setup ourselves; the steps come from T3 Code's docs.
• Julius also teased "cross env orchestration" as coming soon. That is not shipped yet.