Skip to main content
Glama

Claudex

Codex agents for Claude Code.

CI MIT license

Give Claude Code a crew of Codex agents, then follow their work in a local dashboard. Read conversations, inspect changes, steer a running agent, and keep the caller's instructions and context visible in one place.

Claudex is an independent community project. It is not an official OpenAI or Anthropic product.

Claude delegates a review to Codex, then queues and steers a follow-up in the live dashboard

Claude implements CSV export while Codex reviews edge cases. This 28-second excerpt pauses on the Codex launch and task receipt before switching to the live dashboard. Longer feature walkthrough (hosted as a release download, outside Git history). Capture details.

What you get

  • One workspace for your agents. Live activity, searchable conversations, project groups, pinned sessions, and saved history.

  • Control while work is happening. Queue a follow-up, steer the current turn, stop an agent and its delegated children, or resume saved work.

  • A reviewable handoff. Read messages and tool output, inspect per-run unified or split diffs, and preview supported attachments.

  • Context you can inspect. See inherited instructions, memories, skills, model settings, and the runtime's available integrations.

  • One gateway for every caller. The dashboard, CLI, MCP clients, and agent delegation tools share the same local state and delivery rules.

Related MCP server: AgentGo

Quick start

You need Python 3.11+, an installed and authenticated Codex CLI with app-server support, and Claude Code if you want Claude to delegate. Codex access and model usage are provided by your own Codex installation and account.

Install from GitHub with pipx:

pipx install git+https://github.com/danielasher115/claudex.git

There is no PyPI release to install. The repository includes the built dashboard, so Node.js is only needed for UI development.

Windows PowerShell

Start Codex in one terminal and leave it running:

codex app-server --listen ws://127.0.0.1:4500

In a second terminal:

claudex start --codex-url ws://127.0.0.1:4500
claudex status

Linux and WSL

If your Codex installation already provides its local control socket, use it directly:

claudex start
claudex status

For a standalone app-server, start your own Unix socket in another terminal:

codex app-server --listen unix:///tmp/claudex-codex.sock

Then connect to it:

claudex start --socket /tmp/claudex-codex.sock

Open codex.localhost:8765. The public local address stays stable while the backend selects a free port. claudex status prints the current address and connection details. Stop the gateway with claudex stop.

Linux/WSL and native Windows are validated by the test suite, with Python 3.11 and 3.13 checked in CI. macOS is untested. See setup and troubleshooting for platform details, custom ports, and data locations.

Connect Claude Code

Keep Claudex running and add this stdio MCP server to your Claude Code configuration:

{
  "mcpServers": {
    "codex": {
      "command": "claudex",
      "args": ["mcp"]
    }
  }
}

The server name codex matches the optional caller-context hook. Use the same user and data directory as the dashboard. The app's Connect Claude dialog also shows the configuration for your running instance.

For automatic export of the launching Claude conversation, install the context hook:

claudex install-claude-context

This updates your Claude user settings and preserves a backup. Restart or resume Claude Code and reconnect its Codex MCP server. Without the hook, agents can still launch; you can pass a session_summary for the handoff. For Claude on Windows talking to Claudex inside WSL, see the WSL MCP configuration.

See the work

Follow active agents and browse projects:

Claudex overview with active agents and project navigation

Review the agent's recorded changes in a split diff:

Dark split diff review with changed files and line numbers

These two screenshots render the real frontend with fictional local data. Regenerate them without starting Codex or making model requests.

Use it from the CLI

claudex models
claudex call agents.list --json -

For the second command, supply a JSON object on stdin, such as {}. For a launch, save the request below as request.json, replace the workspace path with your project, and run claudex call agents.launch --json request.json:

{
  "prompt": "Review the project and report the most useful small improvement.",
  "cwd": "/absolute/path/to/project",
  "name": "Project review",
  "sandbox": "workspace-write"
}

The launch returns a thread ID immediately. Use agents.status, agents.wait, and agents.result to follow it. See the agent protocol for attachments, exact message targeting, retries, and child agents.

Local by design

Claudex listens on loopback and uses your local Codex app-server. It stores gateway state, prepared attachments, and a local protocol token in your user state directory. Set CLAUDEX_DATA_DIR or pass --data-dir to select another instance. Claudex does not include Codex credentials or host a model service.

Treat the dashboard and MCP connection as access to your local coding agents. Keep them private. Context inheritance can include project instructions, memories, and public caller history. Native Codex sandbox and approval policy remain authoritative. Read security before changing how the gateway is exposed.

Development and documentation

git clone https://github.com/danielasher115/claudex.git
cd claudex
python -m venv .venv
# Activate .venv for your shell, then:
python -m pip install -e '.[dev]'

UI dependencies are installed separately: run npm ci in ui/, then npm run build to rebuild the packaged dashboard. See development for checks and demo screenshots.

Licensed under the MIT License.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables ChatGPT conversations to directly inspect local workspaces and delegate coding tasks to a local Codex agent, with controlled concurrency, reusable sessions, and task/usage tracking.
      603 npm
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables an AI assistant to delegate coding tasks to local Codex or Claude Code CLIs, monitor their progress, read results, continue sessions, and browse the files in user-approved project workspaces. It exposes these capabilities over Streamable HTTP through a Cloudflare tunnel or locally over stdio, so a remote or cloud client can drive agents running on the user's own machine.
      76 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables MCP clients such as Claude Code and Codex to delegate persistent review or isolated write-mode implementation tasks to locally installed Claude Code or Codex runtimes, then inspect progress, approve or deny individual permission requests, and continue the same native conversation. It also lists live model catalogues, lets callers follow up on saved conversations, and can run independently via its bundled CLI.
      2
      MIT