chatgpt-mcp

Local MCP server that drives the ChatGPT web UI via patchright (a stealth Playwright fork). Keeps a persistent Chrome profile so you log in once, then Claude Code, your scripts, and a CLI can share the same ChatGPT session concurrently over CDP.

Exposes three surfaces on top of the same browser controller:

• MCP server (stdio) — for Claude Code and any MCP client.
• HTTP API (localhost, bearer token) — for shell scripts and remote helpers.
• CLI — chatgpt-mcp <subcommand>.

Features

• Persistent login — one-time sign-in in a dedicated Chrome profile; session survives restarts
• Shared session over CDP — launcher keeps Chrome open; MCP server, HTTP API, and CLI attach to the same browser via http://127.0.0.1:9222
• Single-threaded controller — every mutating op goes through a FIFO mutex so concurrent calls never collide
• Long responses — no internal timeout on generation; a single reply can take up to an hour (deep research, Pro Thinking) without anything bailing
• Model switching — Instant / Thinking / Pro via stable data-testid map
• Thinking-level switching — Standard / Longer on Pro/Thinking models, matched by SVG sprite icon (locale-independent) with localized-text fallback
• Structured status — reads current model + thinking level passively from the composer pill (no menus opened)
• Selector self-heal — chatgpt-mcp check walks selectors.json against the live page and reports misses when ChatGPT rotates markup

Quick Start

Requires Node.js 18+ and a local Google Chrome install.

npm install -g @guilhermesilveira/chatgpt-mcp   # installs `chatgpt-mcp` globally

# Terminal 1 — launch Chrome, log in once
chatgpt-mcp launch

# Terminal 2 — use it
chatgpt-mcp status
chatgpt-mcp query "what is 2+2? one word"

How it works

         ┌─────────────────┐    CDP (127.0.0.1:9222)
         │  launcher.mjs   │◀──────────────┐
         │  (persistent    │               │
         │   Chrome tab)   │               │
         └─────────────────┘               │
                                           │
                          ┌────────────────┴──────────────┐
                          │    browser-controller.mjs     │
                          │    (patchright, FIFO mutex)   │
                          └──┬────────────┬────────────┬──┘
                             │            │            │
                     mcp-server.mjs   http-api.mjs   cli.mjs
                        (stdio)      (bearer token)

The first process to call the controller tries to attach over CDP; if the launcher isn't running, it falls back to launching its own persistent context. Either way the profile lives at ~/.chatgpt-mcp/profile/.

CLI

chatgpt-mcp launch              # start Chrome with persistent profile + CDP
chatgpt-mcp server              # run the MCP stdio server (for Claude Code)
chatgpt-mcp http                # run the localhost HTTP API
chatgpt-mcp status              # state=ready model=Pro thinking=Länger
chatgpt-mcp query "prompt..."   # send a prompt, print the reply
chatgpt-mcp last                # print the last assistant message
chatgpt-mcp new                 # open a new chat
chatgpt-mcp model               # print current model
chatgpt-mcp model pro           # switch model
chatgpt-mcp thinking            # print current thinking level
chatgpt-mcp thinking longer     # switch thinking level
chatgpt-mcp stop                # abort an in-flight generation
chatgpt-mcp check               # selector self-heal report

query flags:

chatgpt-mcp query --fresh --model pro --thinking longer "complex question..."

• --fresh — start a new chat first.
• --model <name> — switch model first. Known names: instant, thinking, pro (see selectors.json → model.name_map).
• --thinking <level> — set thinking level first. Known: standard, longer.

MCP tools

Registered by mcp-server.mjs:

| Tool | Description | |------|-------------| | query | Send prompt, wait for full reply, return text. Supports fresh, model, thinking. | | read_last_response | Read the last assistant message without sending anything. | | status | JSON { state, model, thinking }. | | new_chat | Open a new chat. | | set_model / get_model | Switch / read current model. | | set_thinking / get_thinking | Switch / read current thinking level. | | stop | Abort an in-flight generation. |

Wiring it into Claude Code

claude mcp add chatgpt --scope user -- chatgpt-mcp server

(Adjust the command if you did not npm link; use the absolute path to cli.mjs server instead.)

HTTP API

chatgpt-mcp http                # listens on 127.0.0.1:8765

Token is auto-generated at first run and stored at ~/.chatgpt-mcp/token (mode 0600). Pass it as Authorization: Bearer <token>.

TOKEN=$(cat ~/.chatgpt-mcp/token)

curl -s -H "Authorization: Bearer $TOKEN" http://127.0.0.1:8765/status
curl -s -H "Authorization: Bearer $TOKEN" \
  -X POST -H "content-type: application/json" \
  -d '{"prompt":"hello","fresh":true,"model":"pro","thinking":"longer"}' \
  http://127.0.0.1:8765/query

curl -s -H "Authorization: Bearer $TOKEN" http://127.0.0.1:8765/last
curl -s -H "Authorization: Bearer $TOKEN" -X POST http://127.0.0.1:8765/new
curl -s -H "Authorization: Bearer $TOKEN" http://127.0.0.1:8765/model
curl -s -H "Authorization: Bearer $TOKEN" -X POST -d '{"name":"pro"}' http://127.0.0.1:8765/model
curl -s -H "Authorization: Bearer $TOKEN" http://127.0.0.1:8765/thinking
curl -s -H "Authorization: Bearer $TOKEN" -X POST -d '{"level":"longer"}' http://127.0.0.1:8765/thinking
curl -s -H "Authorization: Bearer $TOKEN" -X POST http://127.0.0.1:8765/stop

Request timeout on the server is disabled (requestTimeout=0) so a 1-hour Pro reply won't be cut off.

Selector self-heal

ChatGPT's DOM changes often. If a tool starts failing:

chatgpt-mcp check

Walks every selector in selectors.json against the live page and prints OK/MISS per entry. Exits non-zero if anything is missing. Re-capture the relevant state and update selectors.json directly — no code changes needed.

To re-capture a state:

1. In Chrome DevTools Console: copy(document.documentElement.outerHTML)
2. pbpaste > dom-samples/<state>.html
3. Inspect the HTML and update selectors.json.

For menus (Radix menus close on blur), enable "Emulate a focused page" in DevTools (Cmd+Shift+P → focused) or use the setTimeout trick:

setTimeout(() => copy(document.documentElement.outerHTML), 5000)

then open the menu within 5 seconds.

Tests

Pure-logic unit tests with Node's built-in test runner — no browser, no extra deps:

npm test

Covers the pill-text parser (parse-pill.mjs), CLI flag parser (flags.mjs), and selectors.json shape invariants.

Optional PM2 service

If you want chatgpt-mcp launch to run as a service and survive reboots:

npm install -g pm2
pm2 start chatgpt-mcp --name chatgpt-mcp -- launch
pm2 save
pm2 startup

PM2 is optional; you can also run chatgpt-mcp launch by hand or from a login script. Same recipe works for the HTTP API:

pm2 start chatgpt-mcp --name chatgpt-mcp-http -- http

Note that launch opens a visible Chrome window. On a headless server, that's not going to work — this project is built for local use with a real user session.

Running the MCP server directly

For Claude Code users who haven't npm linked:

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": ["/absolute/path/to/chatgpt-mcp/cli.mjs", "server"]
    }
  }
}

Configuration

All state lives in ~/.chatgpt-mcp/:

| Path | What | |------|------| | ~/.chatgpt-mcp/profile/ | Chrome user data dir (login cookies, settings) | | ~/.chatgpt-mcp/cdp | Current CDP URL (written by the launcher) | | ~/.chatgpt-mcp/token | Bearer token for the HTTP API (mode 0600) |

Environment variables:

| Var | Default | Notes | |-----|---------|-------| | CDP_PORT | 9222 | Port the launcher exposes for CDP | | CHATGPTPRO_CDP | (read from ~/.chatgpt-mcp/cdp) | Override CDP endpoint for the controller | | PORT | 8765 | HTTP API port |

Deployment

There's no "deploy" target — this is a local service that drives a logged-in ChatGPT tab on your machine. If you need it on another machine, clone the repo there and run chatgpt-mcp launch + sign in there. The profile is not portable between machines.

License

MIT