Codex Delegate
Stop burning your frontier agent's limits on boilerplate.
Delegate implementation to the OpenAI Codex CLI — your agent writes the brief and reviews the diff.
[npm version]
[npm downloads]
[MCP Registry]
[codex-delegate-mcp MCP server]
[node]
[license: MIT]
[tests]
[Codex Delegate logo]
Use your best coding agent where its judgment matters most: understanding the task, shaping the plan, and reviewing the result.
Codex Delegate is the MCP bridge that lets Claude Code, Cursor, Copilot — or any MCP client — hand implementation to the OpenAI Codex CLI, then get a clean, structured result back for review.
[A terminal recording of Claude Code delegating to Codex: a health check confirms the Codex CLI, auth and model catalog; an ask-mode run diagnoses why the demo repo prints "today 0", naming commitDay() bucketing by UTC while cellDay() labels by local time, with line numbers; one sentence then fans out to two Codex models at once — one editing the repo, one researching in a separate directory — and the fix lands with tests passing and the footer reading today 4, current streak 23, longest streak 23; finally Codex reviews its own diff and reports three findings]
🧠 Frontier quality, kept
Your assistant does what frontier models are actually for: understands the task, writes a precise brief, reviews the finished diff. Codex holds its own as the implementer — guided and checked by a smarter orchestrator. The result reads like frontier work, because a frontier model planned it and signed off on it.
⚡ Done faster
Codex tears through multi-file edits while a frontier chat model would still be streaming the first file. You delegate, keep working with your assistant, and the diff shows up done.
🔋 Your limits stop being the bottleneck
Delegated work runs on the OpenAI Codex CLI and its own usage — separate from your orchestrator's chat quota, though the Codex side still bills its own way. Your Claude, Cursor, or Copilot subscription spends tokens on the brief and the review; Codex does the grinding. On API? That's the per-token grind moved off your main bill.
[You and your agent understand the task, write the brief and review the diff; the MCP delegate tool hands that brief to the OpenAI Codex CLI, which implements it and edits your workspace; one compact JSON result comes back with what changed, which files, and the thread id]
[A delegate result: one compact JSON block with the final answer, status, thread and delegation ids, workspace, Codex CLI version, cumulative thread token usage, and the files the edit tools reported changing]
Features
- 📦 One result you can review — a compact JSON block: the final answer,
status, the files Codex's edit tools reported changing, cumulative thread token counts, and the threadId to continue from. Fields that carry no signal are omitted.
- 📋 Plan first, then build it on the same thread —
plan returns schema-validated steps for you to approve, and agent implements them. ask answers questions. review runs Codex's own reviewer over uncommitted work, a base branch, or a single commit.
- 🧵 Resume — continue a Codex thread with
resumeThreadId. resumed: false tells you the context did not carry over.
- 🧑🤝🧑 Run several, cancel cleanly — the same question across models, or independent workers on independent directories.
cancel waits for the exit and warns when a process outlives the kill deadline.
- 🤝 One-command install — Claude Code, Google Antigravity, GitHub Copilot CLI, Cursor and VS Code take it as a plugin, with a skill that teaches your agent how to delegate well. Cursor and VS Code install the server in one click; JetBrains, Devin Desktop (Windsurf), Visual Studio and the rest add the stdio server in settings.
- 🩺
doctor — tells you exactly what's missing if setup isn't right.
Install
You need Node.js 20+ and the OpenAI Codex CLI, already logged in (codex login).
Claude Code
Then just ask:
Delegate to Codex: migrate src/api from callbacks to async/await and update the tests, then walk me through what changed.
That's the whole loop — Claude writes the brief, Codex grinds through the files, Claude walks you through the diff.
Cursor
[Add to Cursor]
Or add it under Customize → MCPs in the sidebar, or in ~/.cursor/mcp.json (every project) or .cursor/mcp.json (this project):
Then ask Cursor to delegate implementation to Codex the same way.
Google Antigravity
Restart the Antigravity app to load it. If you added codex-delegate to mcp_config.json by hand, remove that entry.
Without the CLI, use the manual config under More clients.
GitHub Copilot CLI
The server alone, without the skill: copilot mcp add codex-delegate -- npx -y codex-delegate-mcp.
More clients
VS Code — one-click install, or .vscode/mcp.json
[Install in VS Code]
[Install in VS Code Insiders]
Or run Chat: Install Plugin From Source with this repository's URL (needs the chat.plugins.enabled setting).
JetBrains AI Assistant — Settings → Tools → AI Assistant → MCP
Under Settings → Tools → AI Assistant → Model Context Protocol (MCP), click Add and paste this JSON:
Devin Desktop (formerly Windsurf) — ~/.config/devin/mcp_config.json
This config is for Devin Local, the default agent (Local 3.6+). On Windows the file is %APPDATA%\devin\mcp_config.json. Legacy Cascade and older Windsurf builds read ~/.codeium/windsurf/mcp_config.json.
Heads-up: Devin Desktop's Cascade agent caps you at 100 tools across all servers.
Visual Studio 2026 / 2022 — %USERPROFILE%\.mcp.json
Visual Studio 2022 needs 17.14+. Note the top-level key is servers, not mcpServers. A .mcp.json in the solution directory scopes it to that solution.
OpenCode — ~/.config/opencode/opencode.json or project opencode.json
OpenCode does not use mcpServers. Local servers go under mcp, with type: "local" and command as one array:
This is the v1 shape, which OpenCode v2 still reads (v2's native form is mcp.servers).
Google Antigravity (without the plugin) — ~/.gemini/config/mcp_config.json or workspace .agents/mcp_config.json
Kilo Code — kilo.jsonc (mcp key, not mcpServers)
Same shape as OpenCode: type: "local" and command as one array.
In the VS Code extension: Settings → MCP → Add Server → Local (stdio). On Windows, if npx is not found, use command cmd with arguments /c, npx, -y, codex-delegate-mcp.
Zed — Settings → AI → MCP Servers, or context_servers in Zed settings
Zed's native agent uses this, and Zed forwards these servers to external agents over ACP; those agents may also read their own MCP config.
Kiro and any other MCP client
Add the following server to the client's MCP config — for Kiro, ~/.kiro/settings/mcp.json (user) or .kiro/settings/mcp.json (workspace):
MIT © Andrei Lungeanu
Configuration · Security · Privacy · Terms · Changelog
