Desktop Coworker for ChatGPT

Self-hosted MCP coding server and unified local Control Center for ChatGPT web.

Independent community project; not affiliated with or endorsed by OpenAI.

Desktop Coworker for ChatGPT can read and change files, run commands, use Git, restore edit checkpoints, and proxy optional upstream MCP tools. This is high-impact access: read SECURITY.md before connecting it to a remote service.

Windows quick start

Requirements: Windows 10/11 and Git. Everything else is bootstrapped when needed.

git clone https://github.com/colin-dmme/desktop-coworker-for-chatgpt.git
cd desktop-coworker-for-chatgpt
.\run.bat

run.bat is the normal entrypoint. It installs or updates Node.js, installs the locked npm dependencies, builds the TypeScript sources, downloads the official OpenAI Secure MCP Tunnel client when missing, and opens the local Control Center. The first launch requires network access; later launches reuse installed files.

Related MCP server: ChatGPT2LocalBridge

Connect an account

1. In ChatGPT, enable Developer mode under Settings → Security and login. Availability may depend on your account and workspace policy.

2. In OpenAI Platform tunnel settings, create a Secure MCP Tunnel, associate it with the ChatGPT workspace that will use it, and create a Runtime API key. Running and selecting the tunnel requires Tunnels Read + Use permission.

3. Run run.bat; the Control Center opens at http://127.0.0.1:2999/ui/.

4. Choose Add account and enter a workspace, tunnel ID, Runtime API key, and ports.

5. Start the account and wait until MCP, Admin, Tunnel, and Ready are green.

6. Open ChatGPT Plugins, select the plus button, choose Tunnel under Connection, and select the tunnel or paste its tunnel_id.

7. Review the discovered tools, create the connection, and add it to a new conversation from the tools menu. Refresh the connection after server metadata changes.

See the official OpenAI documentation for Secure MCP Tunnel and connecting an MCP server from ChatGPT.

Profiles created by the former PySide dashboard are migrated automatically. Profiles, telemetry, logs, runtime configuration, and upstream MCP settings live outside the repository at %LOCALAPPDATA%\DesktopCoworkerForChatGPT.

The Control Center never returns a Runtime API key to browser JavaScript. The generated tunnel YAML also contains no key; it is passed only to the tunnel child process. The local profile file must still be treated as a credential file.

Conversation access control

Conversation access protection is off by default, preserving the original one-click experience: ChatGPT can use tools as soon as the connector is ready. Owners of shared accounts can opt in per account on the Chat access page. When enabled, a new ChatGPT conversation cannot call computer tools until the local owner approves its pending request or generates a short-lived one-time code. Grants are scoped to one account and workspace, have a read, edit, or execute tier, expire automatically, and can be revoked immediately.

The Chat access history keeps a local audit trail of pending requests, approvals, denials, revocations, protection changes, and one-time-code outcomes. It can be filtered by event, permission, date, fingerprint, label, or tool, and entries can be deleted individually or in filtered batches.

ChatGPT conversation, subject, and organization identifiers are stored only as keyed fingerprints. One-time codes are stored only as keyed hashes, are single-use, expire after two minutes by default, and trigger a temporary lockout after repeated failed attempts. Tool activity recursively redacts password, code, token, authorization, credential, and cookie fields before persistence or streaming.

One interface, three local services

Browser Control Center :2999
  ├─ manages accounts, ports, logs, activity, settings and upstream MCP
  └─ starts one isolated pair per account
       ├─ MCP server :3000 (default)
       │    ├─ filesystem / patches / checkpoints
       │    ├─ shell / processes / Git
       │    ├─ project context / memory
       │    └─ optional upstream MCP servers
       ├─ Admin API :3001 (internal data source; /ui redirects to :2999)
       └─ OpenAI Secure MCP Tunnel health :8080

There is no separate desktop runtime. The web Control Center is localhost-only and is both the launcher UI and the complete administration UI.

Repository layout

public/control/               Unified Control Center frontend
public/control/locales/       Extensible English/Vietnamese UI catalogs
public/ui/                    Legacy per-server fallback UI
src/control/                  Account supervisor, telemetry and Control API
src/admin/                    Per-server localhost Admin API
src/lib/                      MCP services and infrastructure
src/tools/                    MCP tool registration
scripts/windows/              Bootstrap and lifecycle scripts
scripts/development/          Maintainer utilities
config/examples/              Safe example configurations
docs/                         User and architecture documentation
tests/control/                Control Center tests
tests/server/                 MCP server tests
run.bat                       One-click Windows entrypoint

public/ui remains only as a standalone-server compatibility fallback. Normal launches redirect its /ui route to the unified Control Center.

Development

Node.js 22.5 or newer is required because local telemetry uses the built-in SQLite module.

npm ci
npm run build
npm test
npm run test:all
npm run dev

Useful entrypoints:

• npm start — unified Control Center and account supervisor.

• npm run start:server — one standalone MCP server without the supervisor.

• npm run dev — Control Center in TypeScript watch mode.

• npm run dev:server — standalone MCP server in watch mode.

• .\run.bat -SkipLaunch — complete Windows bootstrap smoke test.

The Control Center defaults to English and remembers a language selected in the UI. To add another language, copy a catalog under public/control/locales, keep the same message keys as en.js, and register it in public/control/i18n.js. The i18n test fails when catalogs or referenced keys fall out of sync.

Configuration

The Control Center writes per-account environment files under local app data. For standalone server development, copy .env.example to .env.

The workspace is a default working directory, not a sandbox. Absolute paths and commands run with the current Windows user's permissions.

Documentation

• Control Center guide

• Vietnamese Control Center guide / Hướng dẫn tiếng Việt

• Contributing

• Support

• Code of Conduct

• Changelog

• Security

• Agent onboarding

License

MIT