ChatGPT Web MCP ? local fork

A Node.js MCP server for operating a dedicated signed-in ChatGPT browser session. Fork of Goudu666/chatgpt-web-mcp, based on commit d6c1909. MIT licensed; original attribution and history are retained. Unofficial and dependent on the web interface.

Local additions

• Windows background operation using normal Chrome kept minimized. This is not true headless operation: the tested headless session hit a security check.

• Explicit LOGIN_REQUIRED / HUMAN_VERIFICATION_REQUIRED outcomes and MCP tools to open manual login and minimize the same signed-in browser afterward.

• New conversations routed to an existing CODEX-MCP project, with submission guards and temporary-chat rejection.

• Current Chat/Work button detection and a model/effort picker adapter.

• Work default: GPT-6.1 Sol / Light, verified on the tested account.

• Result-file listing/downloading code with filename sanitization, exclusive saves, bounded waits, and no automatic generation retries.

Related MCP server: ChatGPT Web MCP

Status and known limitations

This is experimental, not a finished integration. Live Chat requests and response extraction worked through the browser functions, including a substantial self-contained design task. This does not establish that every combined MCP tool works on the current UI.

• Chat uses effort controls rather than the Work model selector. The combined preset adapter still needs that distinction implemented; it currently requests a Work model name in Chat and may stop before sending. Do not assume it works.

• The upstream completion detector timed out on a successfully generated document card. The current download tools scope to older response containers and have not been validated on that new card layout. No end-to-end download success is claimed.

• Conversation-type detection can return unknown after navigation.

• Project guards rely on visible UI evidence and URL structure; UI or slug changes may cause safe rejection and need adaptation.

• Model/effort labels and availability vary by account. Browser controls do not independently prove which backend model served a request.

Setup

Requires Node.js 20+, Chrome/Edge, and a ChatGPT account.

npm ci --ignore-scripts
Copy-Item local-project.example.json local-project.json

Edit the ignored local-project.json with your existing CODEX-MCP project URL. The project page is expected to use https://chatgpt.com/g/g-p-<id>/project. No credentials belong in this file. Browser login/session data stays in the separate profile ~/.chatgpt-web-mcp/chrome-profile, never your everyday profile.

Start the server with an absolute path in your MCP client's stdio configuration:

node /absolute/path/to/chatgpt-web-mcp/src/index.js

MCP login: call chatgpt_begin_login, have the user sign in manually, then call chatgpt_finish_login. Login is interactive; normal operation is minimized. If the server reports a security check, stop for manual action. There is no challenge-solving or rate-limit bypass. Pending prompts are never automatically resent by the login tools.

Files and transcripts

chatgpt_list_downloads lists candidate files from the latest response; chatgpt_download_file saves one selected index under ~/.chatgpt-web-mcp/downloads (override with CHATGPT_WEB_DOWNLOAD_DIR). Files are never executed. The 100 MiB limit is checked after transfer, not a bandwidth cap. The newer document-card layout remains an open limitation noted above.

Upstream rotation at 40 turns saves Markdown transcripts locally under ~/.chatgpt-web-mcp/conversation-context before creating a new conversation. Archives are not automatically injected into new chats. Selective retrieval can reduce context, but no archive-search tool is implemented here. Browser profiles, project IDs, runtime state, downloads and private transcripts must stay out of Git.

Validation

npm test

Run tests after creating the local project configuration. Focused local checks cover defaults, safe file saves, project guards and login recovery; they do not replace live UI verification. No automatic account rotation or generation retries are provided. Ask for user approval before delegating work or uploading files.

The original upstream English README is preserved in docs/UPSTREAM-README.en.md.