Codex As MCP Server

@kky42

Basic information

About Codex As MCP Server

Enable Claude Code, Cursor and other AI tools to call Codex for task execution. Plus/Pro/Team subscribers can maximize GPT-5 usage without additional costs.

#gpt-5#codex #openai #claude-code #claude

Codex As MCP Server config

Add this server to your MCP-compatible client using the configuration below.

{
  "mcpServers": {
    "codex": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "codex-as-mcp@latest",
        "--yolo"
      ],
      "env": {}
    }
  }
}

Codex As MCP Server tools

Execute prompt using codex for general purpose. Args: prompt (str): The prompt for codex work_dir (str): The working directory, e.g. /Users/kevin/Projects/demo_project ctx (Context): MCP context for logging

Execute code review using codex with pre-defined review prompts for different scenarios. This tool provides specialized code review capabilities for various development scenarios, combining pre-defined review templates with custom instructions. Args: review_type (str): Type of code review to perform. Must be one of: - "files": Review specific files for code quality, bugs, and best practices Target: comma-separated file paths (e.g., "src/main.py,src/utils.py") Example: review_type="files", target="src/auth.py,src/db.py" - "staged": Review staged changes (git diff --cached) ready for commit Target: not needed (automatically detects staged changes) Example: review_type="staged" - "unstaged": Review unstaged changes (git diff) in working directory Target: not needed (automatically detects unstaged changes) Example: review_type="unstaged" - "changes": Review specific commit range or git changes Target: git commit range (e.g., "HEAD~3..HEAD", "main..feature-branch") Example: review_type="changes", target="HEAD~2..HEAD" - "pr": Review pull request changes comprehensively Target: pull request number or identifier Example: review_type="pr", target="123" - "general": General codebase review for architecture and quality Target: optional, can specify scope or leave empty for full codebase Example: review_type="general", target="src/" work_dir (str): The working directory path (e.g., "/Users/kevin/Projects/demo_project") target (str, optional): Target specification based on review_type: - For "files": comma-separated file paths - For "staged"/"unstaged": not needed (leave empty) - For "changes": git commit range (commit1..commit2) - For "pr": pull request number/identifier - For "general": optional scope (directory path or leave empty) prompt (str, optional): Additional custom instructions to append to the review prompt. Use this to specify particular aspects to focus on or additional context. Example: "Focus on security vulnerabilities and performance" ctx (Context, optional): MCP context for logging Returns: str: Detailed code review results from codex Examples: # Review specific files with security focus codex_review("files", "/path/to/project", "src/auth.py,src/api.py", "Focus on security vulnerabilities") # Review staged changes before commit codex_review("staged", "/path/to/project") # Review unstaged work-in-progress changes codex_review("unstaged", "/path/to/project", "", "Check for incomplete implementations") # Review recent commits codex_review("changes", "/path/to/project", "HEAD~3..HEAD", "Look for performance regressions") # Review pull request codex_review("pr", "/path/to/project", "456", "Focus on test coverage") # General codebase review codex_review("general", "/path/to/project", "src/", "Identify technical debt")

Overview

What is Codex As Mcp?

Codex As Mcp is an MCP server that enables AI tools such as Claude Code and Cursor to call OpenAI Codex for task execution. It is designed for Plus, Pro, and Team subscribers who want to maximize GPT‑5 usage without incurring additional costs.

How to use Codex As Mcp?

First install the Codex CLI globally via npm and log in. Then add the server to your MCP configuration (.mcp.json) using uvx codex-as-mcp@latest. For writable (full‑capability) mode include the --yolo flag. Alternatively, use claude mcp add commands from the terminal.

Key features of Codex As Mcp

Exposes two tools: codex_execute and codex_review
Safe Mode (read‑only) and Writable Mode (--yolo)
Sequential execution prevents agent conflicts
Works with Claude Code, Cursor, and other AI tools
No extra costs for eligible Codex subscribers

Use cases of Codex As Mcp

Run general‑purpose code tasks via an AI assistant
Perform specialized code reviews with targeted prompts
Integrate Codex’s capabilities into agentic workflows

Frequently asked questions

Which AI tools can use this server?

Claude Code, Cursor, and any other tool that supports the Model Context Protocol (MCP) can connect to Codex As Mcp.

What is the difference between Safe Mode and Writable Mode?

Safe Mode (default) restricts operations to read‑only for safety. Writable Mode (activated with `--yolo`) allows full Codex capabilities, including file modifications.

Do I need a Codex subscription?

Yes, this server is intended for Codex Plus, Pro, or Team subscribers. Using it can help maximize GPT‑5 usage without additional costs.

What runtime dependencies are required?

You must have Node.js installed to run `npm install -g @openai/codex`. The server itself is launched via `uvx`, which requires a compatible Python environment.

Can I create custom use cases beyond the provided tools?

The README invites users to open an issue if they have other use‑case requirements, implying the two built‑in tools can be extended upon request.

Comments

More Developer Tools MCP servers

ArtiBot Cloud MCP - Zhijiangyun 智匠云

bofob

ArtiBot Cloud MCP (Zhijiangyun 智匠云) — cloud MCP server for robot manipulation data & GPU workloads. 15 tools: robot fleet management, data collection & video recording, post-processing pipelines, and benchmark evaluation

Vibgrate MCP

Vibgrate

Query your team's drift, vulnerability, and migration data from any AI assistant. Vibgrate MCP connects Cursor, Claude, ChatGPT, Windsurf, or VS Code to Vibgrate Cloud: 51 tools for DriftScores, CVEs and EOL runtimes, up

Casefile

azimov777

One case file per task — shared by every session, agent and person that touches it. Self-hosted MCP server and live board. MIT.

Jaybird Digital

Steve Wheeler

Jaybird Digital lets you build a website for you in the conversation. Tell your ai about your business and what you want, and it writes the pages and saves them to Jaybird. You get a link to view the site right away, wi

SSH MCP Server

hypnosis

SSH MCP server for AI agents: remote commands, file transfer, log search and server audits through OpenSSH.