Codex Specialized Subagents
MCP server that lets Codex delegate to isolated codex exec sub-agents, selecting repo+global skills automatically
Install
git clone https://github.com/leonardsellem/codex-specialized-subagentsInstall command taken from the project's README or registry entry. Always review code before running it.
README excerpt
Read full README on GitHub →codex-specialized-subagents
Artifact-first sub-agent delegation for Codex CLI (MCP server).
This repo provides a local (stdio) MCP server that exposes:
delegate_autopilot— decide whether to delegate and, if yes, orchestrate one or morecodex execsub-agent runsdelegate_run— run a single specialist sub-agent viacodex execdelegate_resume— resume a prior sub-agent thread viacodex exec resume
Each tool call writes a run directory under ${CODEX_HOME:-$HOME/.codex}/delegator/runs/<run_id>/ containing the prompt, selected skills, event stream, and structured results (artifact-first debugging).
When to use
- You want parallelism and specialization for multi-step / cross-cutting work.
- You want durable artifacts (logs + outputs) to debug and review what happened.
Requirements
- Node.js
>=20(seepackage.json#engines) npmmise(recommended): installs the pinned runtime frommise.tomlcodexCLI on your PATH and authenticated (required for real delegation runs)
Optional:
- Python 3 (only for helper scripts under
.agent/)
Install & quickstart (from source)
From the repo root (installs deps + builds dist/):
# Recommended: install the pinned runtime (see mise.toml)
mise install
# Drift check (lockfiles + pins)
./toolchain-check.sh
# Install deps from lockfile
npm ci
# Build
npm run build
Configure Codex (recommended, prevents timeouts)
Delegated runs can take minutes. Set this server’s MCP tool timeout to 1200 seconds (20 minutes) in your Codex config ($HOME/.codex/config.toml):
mkdir -p "$HOME/.codex"
cat >> "$HOME/.codex/config.toml" <<'EOF'
[mcp_servers.codex-specialized-subagents]
tool_timeout_sec = 1200
EOF
If you already have a [mcp_servers.codex-specialized-subagents] section, edit the existing tool_timeout_sec instead of appending a duplicate.
Common gotcha: tool_timeout_sec is not an env var. If you put it under mcp_servers.codex-specialized-subagents.env.*, Codex will error with “expected a string” (env values must be strings).
If you see a TOML parse error about a “duplicate key” for [mcp_servers.codex-specialized-subagents], you have that table declared twice — keep only one header and put tool_timeout_sec = 1200 inside it.
Register with Codex (recommended defaults)
From the repo root (includes per-job reasoning-effort overrides for delegate_autopilot):
codex mcp add codex-specialized-subagents \
--env CODEX_AUTOPILOT_REASONING_EFFORT_LOW=low \
--env CODEX_AUTOPILOT_REASONING_EFFORT_MEDIUM=medium \
--env CODEX_AUTOPILOT_REASONING_EFFORT_HIGH=high \
-- node "$(pwd)/dist/cli.js"
Verify:
codex mcp get codex-specialized-subagents
Remove:
codex mcp remove codex-specialized-subagents
Usage
Interactive autopilot (recommended)
In Codex interactive mode, delegate_autopilot can split a request into jobs (scan / implement / verify) and run specialist sub-agents.
To make delegation feel automatic in interactive mode, install the included delegation-autopilot skill globally:
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills/delegation-autopilot"
cp .codex/skills/delegation-autopilot/SKILL.md \
"${CODEX_HOME:-$HOME/.codex}/skills/delegation-autopilot/SKILL.md"
Then try prompts like:
- “Refactor the MCP server and update tests + README.”
- “Audit the repo docs and propose improvements.”
Similar plugs
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
An AI skill that provides design intelligence for building professional UI/UX across multiple platforms.
Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local…
Why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
Production-grade engineering skills for AI coding agents.