Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@codex-claude-bridge-mcp Create a task to review the PR and list all open tasks"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
English · Türkçe
One local MCP server that lets Codex and Claude Code coordinate through a shared task board and message inbox.
It uses the standard MCP stdio transport. No API key, web server, or cloud account is required. Both clients run the same server program and point at the same local data directory.
Scope: version 0.1.0 is for a single machine or a trusted shared filesystem. It is not a secure multi-user or internet-facing collaboration service.
Create, list, claim, update, block, and complete shared tasks.
Send general messages or messages attached to a task.
Prevent active tasks from being claimed by both agents.
Atomically write the shared state, with a cross-process lock for concurrent client calls.
Related MCP server: session-coord-mcp
git clone https://github.com/nevzoom/codex-claude-bridge-mcp.git
cd codex-claude-bridge-mcp
npm install
npm run build
.\scripts\install.ps1This registers the server with both locally installed CLIs and stores shared state in ~/.codex-claude-bridge. Restart Codex and Claude Code afterwards.
To register only one client:
.\scripts\install.ps1 -Codex
.\scripts\install.ps1 -Claudegit clone https://github.com/nevzoom/codex-claude-bridge-mcp.git
cd codex-claude-bridge-mcp
npm install
npm run build
bash scripts/install.shUse bash scripts/install.sh --codex or --claude to register only that client. To use another shared directory, pass --data-dir /path/to/bridge-data (or -DataDirectory in PowerShell).
The installer runs the official CLIs for you. If you prefer to configure manually, register this stdio command with both clients, using the same data directory:
node /absolute/path/to/codex-claude-bridge-mcp/dist/index.jsSet this environment variable for both registrations:
CODEX_CLAUDE_BRIDGE_DATA_DIR=/absolute/path/to/shared-bridge-dataCodex also supports project/global MCP configuration in config.toml; Claude Code supports project/user MCP registration. The supplied installers use user-level registration so the bridge is available across projects.
When this variable is omitted, the server uses ~/.codex-claude-bridge. This is convenient when both clients run under the same operating-system user. Set it explicitly whenever clients run under different users or machines.
Tool | Purpose |
| Add a shared task. |
| View tasks, optionally by status. |
| Mark a task as owned by Codex, Claude, or a human. |
| Change its status, owner, or result. |
| Send a general or task-linked message. |
| Read messages for an agent or task. |
Create a discrete task with create_task.
The receiving agent calls list_tasks, then claim_task before working.
Share important context with post_message.
Finish with update_task using status: "done" and a concise result; use blocked if human input is needed.
npm install
npm run check
npm run buildCreate an empty GitHub repository named codex-claude-bridge-mcp, commit this project, and push the main branch.
Create an npm account, then run npm login and npm publish. If the package name is taken, rename the name field to an npm scope you own, such as @your-name/codex-claude-bridge-mcp.
Create a GitHub release and tag it with the same semantic version as package.json.
The package only publishes the compiled dist directory plus the README and MIT license. CI type-checks and builds every pull request.
After publishing to npm, users can register the package directly with their MCP clients using an npx -y <package-name> stdio command. The GitHub-clone installer remains the easiest supported setup because it registers both clients and a shared data directory in one step.
Task text and messages are stored unencrypted in state.json under the chosen data directory. Do not point it to an untrusted network share and do not put secrets into messages. The data directory is intentionally excluded from Git.
Claim an open task before starting work; prevents duplicate ownership.
Inputs: owner (required), task_id (required).
Create a shared task that either Codex or Claude Code can claim.
Inputs: title (required), created_by (required), description.
List tasks in the shared coordination board.
Inputs: status.
Send a message to the other agent, optionally attached to a task.
Inputs: from (required), text (required), to, task_id.
Read shared messages, optionally filtering for a task or recipient.
Inputs: task_id, recipient.
Update a task's status, owner, or final result.
Inputs: task_id (required), owner, result, status.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 6 tools
Each tool targets a distinct action: creating, listing, claiming, updating tasks, and posting/reading messages. There is no overlap or ambiguity between any two tools.
All tool names follow a consistent verb_noun pattern (create_, list_, claim_, update_, post_, read_). The naming is predictable and uniform.
Six tools is a well-scoped size for a coordination bridge. Each tool earns its place and covers a necessary function without bloat.
The tool set covers the full task lifecycle (create, list, claim, update) and messaging (post, read). A minor gap is the absence of a dedicated get_task or delete/archive mechanism, but update_task can handle status changes.
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
Shared memory and mail for your AI agents. Verified with Claude Code; other MCP clients in testing.
Hosted X Ads MCP server for Claude, ChatGPT and Cursor, with each action switched per account.
Official MCP server for Agentwork — delegate tasks to AI agents with human-in-the-loop
