Farcall MCP

Let Claude Code call Codex, or Codex call Claude Code, through one pending MCP call.

The worker builds or reviews. The parent waits for the result, then resumes the same session for corrections. No status polling needed.

Install

In Claude Code, install the Codex worker.

/plugin marketplace add regenrek/farcall-mcp
/plugin install codex-worker@farcall

Start the parent with automatic MCP backgrounding disabled.

CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS=0 MCP_TOOL_TIMEOUT=7200000 claude

In Codex, install the Claude worker.

codex plugin marketplace add regenrek/farcall-mcp
codex plugin add claude-worker@farcall

Start a new session after installation. In Codex 0.147.0 or newer, the plugin keeps the worker tool outside Code Mode without extra configuration. Follow the host setup guide for persistent settings, manual installation & a check that your host waits without polling.

Related MCP server: Code Worker MCP

Use it

Ask Claude Code to use /codex-worker:codex-worker, or Codex to use $claude-worker, with your task. For example,

Use Claude worker with this task: "Add CSV export to the reports page and run the tests."
Give me the changes and test results back.

The run tool accepts the task directly.

{
  "cwd": "/absolute/path/project",
  "delegation_id": "review-001",
  "prompt": "Review the current changes. Do not edit files. Return concrete findings.",
  "model": "claude-opus-5-5",
  "effort": "high"
}

For 1–5 parallel workers in separate checkouts, use one run_batch call. Batch usage, corrections & recovery.

Change model to an identifier supported by the worker CLI. Models, permissions & session resume.

Full logs are opt-in with trace: true. Normal runs keep only the state needed for retries & resume, plus the full answer if the returned preview is shortened. Local state & tracing.

Before a long run, check the worker account & keep returned results short.

Stale locks recover automatically only when the same-host server and all recorded worker groups are gone. Legacy or uncertain locks still require inspection. Before removing a worker checkout, copy its artifacts/farcall/<delegation_id>/ records elsewhere, only after .active is gone. Crash recovery & record preservation.

Polling vs. MCP waiting

Farcall keeps the directly delegated MCP call open until its result. If a worker starts native Codex subagents, Farcall does not control that worker's wait_agent, sleep or status loops. The parent can be waiting while short native timeouts cause extra model turns inside the worker; that alone does not establish a Farcall waiting bug. Host scheduling still matters: an open MCP call does not guarantee that a host never polls. Native wait configuration belongs to the calling workflow, outside Farcall.

Two runs of the same task, with estimated API costs rather than subscription spend. Implementation & review also differed, so polling does not explain the whole difference. Run J used an earlier version of the MCP worker.

See the full experiment, results & cost breakdown.

Development

Run pnpm install --frozen-lockfile & pnpm check. Build generated bundles from src/ with pnpm build.

Architecture · Verification · Release steps