codex-async-mcp

Local MCP server that wraps the codex CLI asynchronously — returns a job_id immediately instead of blocking, so Claude never hits the MCP protocol timeout (-32001).

Requirements

• Python 3.11+
• codex CLI installed and on $PATH (v0.125.0+)
• Claude Code CLI

Installation

cd ~/payroll-mcp   # or wherever this repo lives
pip install -e ".[dev]"

Verify:

python -c "from codex_async_mcp.server import mcp; print(mcp.name)"
# → codex-async-mcp

Register with Claude

Global (all projects)

claude mcp add codex-async -s user -- python -m codex_async_mcp.server

Project-only

cd ~/payrollservice-thailand   # or any project
claude mcp add codex-async -- python -m codex_async_mcp.server

Verify

claude mcp list
# codex-async: python -m codex_async_mcp.server - ✓ Connected

Add tool permissions (settings.local.json)

{
  "permissions": {
    "allow": [
      "mcp__codex-async__codex_start",
      "mcp__codex-async__codex_poll",
      "mcp__codex-async__codex_list",
      "mcp__codex-async__codex_cancel"
    ]
  }
}

Tools

approval_policy values

For Claude automation always use full-auto — suggest mode waits for interactive input which will never arrive inside a subprocess.

Example usage

codex_start(
  prompt="In app/services/prorate_calculation_service.rb line 96, change format(...) to number_to_currency(...)",
  cwd="/Users/bbgummybear/payrollservice-thailand",
  approval_policy="full-auto"
)
# → { job_id: "f3a9b2", status: "running", pid: 12345 }

codex_poll(job_id="f3a9b2")
# → { status: "running", output: "Reading file..." }

codex_poll(job_id="f3a9b2")
# → { status: "done", exit_code: 0, output: "Applied changes to prorate_calculation_service.rb" }

Job state

Jobs are stored in ~/.codex-async/jobs/{job_id}/:

~/.codex-async/jobs/f3a9b2/
  meta.json     ← status, pid, timestamps, exit_code
  output.txt    ← stdout + stderr from codex

meta.json structure:

{
  "job_id": "f3a9b2",
  "status": "running | done | error | cancelled",
  "prompt": "...",
  "cwd": "/path/to/repo",
  "approval_policy": "full-auto",
  "pid": 12345,
  "started_at": "2026-04-29T10:00:00+00:00",
  "finished_at": null,
  "exit_code": null
}

Troubleshooting

codex-async: ... - ✗ Failed in claude mcp list

Python can't be found or the package isn't installed in the right environment.

# Check which python Claude is using
which python

# If using conda, register with the full path
claude mcp add codex-async -s user -- /Users/bbgummybear/miniconda3/bin/python -m codex_async_mcp.server

# Verify the package is installed in that environment
/Users/bbgummybear/miniconda3/bin/python -c "import codex_async_mcp; print('ok')"

status: "error" immediately after codex_start

Codex failed to start. Check the raw output:

cat ~/.codex-async/jobs/<job_id>/output.txt

Common causes:

status: "running" forever, never finishes

The subprocess is hung (waiting for input or stuck in a loop).

# Check if the process is still alive
ps aux | grep codex

# Check live output
tail -f ~/.codex-async/jobs/<job_id>/output.txt

# Cancel the job
codex_cancel(job_id="<job_id>")

Most common cause: using approval_policy="suggest" which pauses for interactive approval. Use "full-auto" instead.

Job shows status: "running" after server restart

The MCP server lost the in-memory Popen registry on restart. The next codex_poll call will detect the PID is dead and update the status automatically.

codex_poll(job_id="<job_id>")
# → { status: "done", ... }   ← auto-resolved on first poll

Old jobs filling up disk

# View all jobs sorted by date
ls -lt ~/.codex-async/jobs/

# Delete jobs older than 7 days
find ~/.codex-async/jobs -maxdepth 1 -type d -mtime +7 -exec rm -rf {} +

Project structure

codex-async-mcp/
├── README.md
├── pyproject.toml
├── src/
│   └── codex_async_mcp/
│       ├── __init__.py
│       ├── server.py        # MCP entry point, tool definitions
│       ├── job_manager.py   # spawn / poll / cancel / list
│       └── config.py        # JOBS_DIR, CODEX_BIN, defaults
└── tests/
    └── test_job_manager.py

Run tests

pytest tests/ -v