codex-mcp-swarm

An MCP server that wraps OpenAI's Codex CLI with true parallel execution and live task monitoring. Zero dependencies -- single Python file, stdlib only.

Why?

The official codex mcp-server processes requests sequentially. If your MCP client (Claude Code, etc.) needs to run 5 Codex tasks, they queue up one after another. This server spawns each task as an independent subprocess, so they run in parallel.

Unique features no other Codex MCP wrapper has:

• Worktree isolation -- worktree: true creates an isolated git worktree per task so parallel Codex instances never edit past each other

• Batch wait -- launch N tasks, call codex_wait once, get all results when they finish

• Live status -- see what each Codex task is doing right now (last tool call, current reasoning, progress)

• Full flag parity -- same parameters as the official Codex MCP tool (sandbox, approval-policy, cwd, model, config, etc.)

• Drop-in config -- accepts the same -c key=value server args as codex mcp-server

Related MCP server: par5-mcp

Tools

Installation

Claude Code

claude mcp add codex-swarm -- uvx --upgrade codex-mcp-swarm \
  -c model=gpt-5.4 \
  -c approval_policy=never \
  -c sandbox_mode=danger-full-access \
  --skip-git-repo-check

That's it. No clone, no setup. uvx downloads and runs it directly from PyPI. The --upgrade flag ensures you always get the latest version on restart.

Note: Requires uv (curl -LsSf https://astral.sh/uv/install.sh | sh). Alternatively, use pipx run codex-mcp-swarm instead of uvx codex-mcp-swarm.

Manual (~/.claude.json)

{
  "mcpServers": {
    "codex-swarm": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "--upgrade",
        "codex-mcp-swarm",
        "-c", "model=gpt-5.4",
        "-c", "approval_policy=never",
        "-c", "sandbox_mode=danger-full-access",
        "--skip-git-repo-check"
      ]
    }
  }
}

The -c flags are identical to codex mcp-server -- copy-paste your existing config.

Usage

Parallel execution

1. Call codex_async with prompt A  -->  task_id: "abc123"
2. Call codex_async with prompt B  -->  task_id: "def456"
3. Call codex_async with prompt C  -->  task_id: "ghi789"
4. Call codex_wait(task_ids=["abc123", "def456", "ghi789"])
   --> blocks until all finish, returns all results

Collecting results: read this before you fan out

A task_id is internal to this server. Your MCP client does not track it. It will not appear in a task list, and no notification fires when the task finishes. If the agent's turn ends before codex_wait is called, the Codex run still completes and writes its result to disk, but nothing is left to deliver it and nothing wakes the session. codex_status is a progress peek only; it collects nothing.

So: codex_async and codex_wait belong in the same turn.

Let the wait run long. Claude Code moves any main-conversation tool call still running after two minutes into a tracked background task and re-invokes the session with the result when it settles -- that is the wake-up mechanism, so crossing the two-minute line is the goal rather than something to dodge. Short timeout values are raised to a floor (default 150s, override with CODEX_SWARM_MIN_WAIT) for exactly this reason. Tasks that have already finished return instantly regardless of the floor.

Worktree isolation

Prevent parallel tasks from editing the same files:

1. Call codex_async(prompt="Refactor auth", worktree=true)
   --> task_id: "abc123"
   --> Worktree Branch: codex-swarm/abc123

2. Call codex_async(prompt="Add logging", worktree=true)
   --> task_id: "def456"
   --> Worktree Branch: codex-swarm/def456

3. codex_wait(task_ids=["abc123", "def456"])
4. git merge codex-swarm/abc123
5. git merge codex-swarm/def456

Each task gets its own git worktree and branch based on HEAD. After completion, merge the branches back. Worktrees are automatically cleaned up after 24 hours (configurable via CODEX_SWARM_TASK_MAX_AGE).

Live monitoring

Call codex_status(task_ids=["abc123"])
-->
=== Task abc123 (45s elapsed) ===
Phase: running
Tools called: 23
Last tool: exec_command(grep -rn "handleError" src/)
Output: Analyzing error handling patterns across the codebase...

Session continuity

1. Call codex(prompt="Review this file")  -->  result + session persisted
2. Call codex_reply(threadId="<session-uuid>", prompt="Now fix the bug you found")

For a follow-up that will run more than a few minutes, resume in the background instead -- codex_async accepts threadId and gives you a task_id that survives a client idle timeout:

1. Call codex_async(threadId="<session-uuid>", prompt="Now implement it")
                                          -->  task_id: "abc123"
2. Call codex_wait(task_ids=["abc123"])   -->  result, with full prior context

codex_reply is synchronous and emits progress notifications while it waits, but a client that gives up on silence will still abandon the request. Only the async path leaves you a handle to recover with.

Server flags

Per-call parameters

All parameters from the official Codex MCP tool are supported:

• prompt (required)

• model -- override server default

• sandbox -- read-only, workspace-write, danger-full-access

• approval-policy -- untrusted, on-failure, on-request, never

• cwd -- working directory

• profile -- config profile from config.toml (ignored on threadId resumes -- codex exec resume has no --profile flag)

• config -- object of key=value overrides

• worktree -- run in an isolated git worktree (prevents parallel tasks from conflicting)

• base-instructions, developer-instructions, compact-prompt

MCP Resources

The server exposes read-only resources for discoverability:

Environment variables

Requirements

• Python 3.8+

• Codex CLI installed and authenticated

• No pip dependencies (stdlib only)

• Works on Linux and macOS (Linux gets extra PID reuse protection and zombie detection via /proc)

Credits

Originally inspired by jeanchristophe13v/codex-mcp-async. Rewritten with full flag parity, JSONL status parsing, batch wait, and session reply support.

License

MIT