Crossfeed Chat turns the ChatGPT subscription you already pay for into a model your local coding agents can call, and lets your own ChatGPT chats use your computer.

Your ChatGPT plan's chat messages are already paid for. Chat has its own usage limits, separate from Codex and from the API (OpenAI's usage guidance, checked 4 October 2026). Crossfeed Chat points an agent such as Pi at that allowance through an OpenAI-style endpoint, so requests that would cost API tokens or Codex allowance run on your plan instead. Your plan's own limits still apply.

It also works the other way round. A second MCP connector gives your ChatGPT chats your local files, Git, guarded shell commands, and selected Gaddi, OmniRead and qmd tools, so a ChatGPT chat can read, search and fix code on your computer.

It runs on macOS and Linux, and works alone or with Crossfeed Orchestrator, which routes work across agents.

How it does at coding

We measured it with the Crossfeed Orchestrator measurement suite, which scores coding agents on fix, implement, review, repository-question, reasoning and extraction tasks at four difficulty levels, with hidden checks. On 12 of its tasks (5 October 2026), ChatGPT through Pi passed 9 of 11 measured tasks at High and 9 of 10 at Extra High, against 12 of 12 for Codex Sol High. Median times were 341 seconds for ChatGPT High and 232 for Sol High. Use Codex when its allowance is there, and ChatGPT when you want a second, separate allowance.

A ChatGPT chat repairs a failing test with the local workbench:

A ChatGPT worker picks up a queued repair request from a local agent:

Account risk and limits

Automating the ChatGPT web app may conflict with OpenAI's terms. The Terms of Use and Europe Terms of Use prohibit automatically or programmatically extracting output and bypassing rate limits, and allow suspension for a breach (checked 4 October 2026). I accept this risk for my account. You decide whether to accept it for yours.

The ChatGPT interface can change and break a wake. Answers arrive whole, not streamed, token usage is unknown, and only a subset of Chat Completions is supported. When you need dependable API behaviour, use the API.

Install

You need Node.js 24.18 or newer, Chrome 120+ or a Chromium browser such as Brave, an HTTPS tunnel (such as Cloudflare Tunnel), and a ChatGPT plan that supports custom MCP connectors. There is no server build step and no runtime dependency: Node runs the TypeScript source directly. The unpacked extension is TypeScript too, so it needs one build before Chrome can load it.

Clone the repository, put a 32-256 character CROSSFEED_API_KEY in the environment from your password manager, and start the server:

node src/cli.ts serve --public-url https://your-tunnel.example --connector crossfeed

Tunnel only 127.0.0.1:4320, the MCP listener; keep the control API on 127.0.0.1:4319. Then register the connector in ChatGPT:

node src/cli.ts link                 # opens registration for five minutes
node src/cli.ts approve --pending    # after adding https://your-tunnel.example/mcp in ChatGPT
node src/cli.ts approve <id>

Run npm ci && npm run build before loading extension/ unpacked.

Load extension/ as an unpacked extension (Developer mode), approve its pairing code, and add a worker:

node src/cli.ts extension approve <code>
node src/cli.ts worker add sol --row "GPT-6" --level 2

sol is a name you choose; --row is the model's label in ChatGPT's model menu and --level its thinking level (0 Instant to 4 Pro). Point your agent at http://127.0.0.1:4319/v1 with your key and the model chatgpt:sol. To give your own chats local tools, start the server with --workbench /path/to/repo and add the crossfeed-tools connector at /tools/mcp.

docs/usage.md covers wake modes, several chats per worker, attachments, deadlines, Pi sessions, the workbench and its shell modes, the tools bridge, Linux setup and stored data.

Checks

npm ci && npm test && npm run typecheck

The tests run the production handlers, OAuth isolation, attachments, persistence, wake controls and command guards against real local listeners. They never contact ChatGPT.

Licence

Apache 2.0, with third-party notices in NOTICE. The name and logo are outside the licence.