Codex Bridge

What Codex Bridge MCP server does

Codex Bridge MCP server links MCP clients and local project workspaces through the Codex Bridge desktop application and background service. It can provide direct access to authorized project directories for file operations, patches, command execution, and Git actions. Those operations follow the registered project, Direct Workspace rules, and approval settings.

The service can also delegate work to connected agent engines. Supported integrations listed by the project include Codex, OpenCode, DeepSeek Harness, Antigravity, Pi, and Qoder. A desktop workbench groups conversations and tasks by agent, shows tool activity and approvals, stores task history, and supports retries and handoffs between agents.

This places Codex Bridge in the Coding Agents category. It fits local development workflows where an external MCP client needs to interact with files or start work on a project without exposing unrestricted access to the machine.

How it works

The desktop app runs a local service that receives MCP requests from configured clients. Qwen Studio can connect through local HTTP MCP. Gemini can use a local MCP command when it runs on the same computer, or a cloud MCP connection through ChatGPT Sites or ngrok. ChatGPT and Dot use the documented OpenAI Secure MCP Tunnel setup.

Projects must be registered before they can be used. Registering a directory authorizes access to that project. Direct operations may require approval, and delegated tasks use the selected project, agent, model preference, and read-only or full permission mode. Full permission includes writing and network access; read-only mode does not allow writing or tool-based network access.

Remote requests may omit a project identifier or provider identifier. In those cases, the service uses the workbench default project or Codex provider respectively. Active sessions provide live output, while task cards show tool execution, approval requests, structured questions, and results.

Setup and configuration

Download a release for macOS, Windows, or Ubuntu Linux, or build from source with the platform-specific Swift and native toolchain requirements. macOS packages target version 14 or later and are ad-hoc signed rather than Apple-notarized. Windows requires WebView2 Runtime.

After launching the app, confirm that its background service is connected, register one or more project directories, and configure an MCP client using the connection method appropriate to that client. Agent connections are optional for direct file and command work. To delegate tasks, connect an installed agent from the Bridge connection page and configure its login or API credentials as required by that provider.

Credentials are stored through the operating system credential store. Remove credentials before sharing logs, configuration, or screenshots.

Tools and capabilities

Codex Bridge MCP server supports the following documented areas:

• Direct file reads and writes, patches, command execution, and Git operations
• Agent task submission through connected clients
• Task approvals, structured questions, live output, interruption, and follow-up instructions
• Workbench sessions, paginated history, retries, tool cards, and cross-agent handoff
• Local skill discovery, read-only inspection, and explicit skill actions
• Model, permission, and session controls for supported agent providers

Writing is limited to one active write task per project across all agents. Codex supports up to four concurrent execution sessions. Other agent concurrency depends on provider limits and local resources; tasks are rejected or fail to start when a required slot is unavailable rather than being automatically queued.

Limitations and notes

Available operations depend on the connected agent, connection detection, and selected task permissions. Agent-specific login and API configuration is separate from Gemini MCP setup. Gemini custom MCP connections work only in Gemini task mode.

The service is designed around authorized local directories and a running Bridge application. External model requests are sent to the service selected by the user. The README does not document a general environment-variable configuration, so deployment should follow the desktop installation and client-specific connection guides rather than assume a standalone server process.