codex-mcp-agent
README.md
# codex-mcp-agent
[中文版](./README.zh-CN.md)
Note: This project is a fork of the current Git remote:
- origin: https://github.com/1WorldCapture/codex-as-mcp
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.
## Setup
### 1. Install Codex CLI
```bash
npm install -g @openai/codex
codex login
```
### 2. Configure MCP
Add to your `.mcp.json`:
**Safe Mode (Default):**
```json
{
"mcpServers": {
"codex": {
"type": "stdio",
"command": "uvx",
"args": ["codex-mcp-agent@latest"]
}
}
}
```
**Writable Mode:**
```json
{
"mcpServers": {
"codex": {
"type": "stdio",
"command": "uvx",
"args": ["codex-mcp-agent@latest", "--yolo"]
}
}
}
```
Or use Claude Code commands:
```bash
# Safe mode (default)
claude mcp add codex-mcp-agent -- uvx codex-mcp-agent@latest
# Writable mode
claude mcp add codex-mcp-agent -- uvx codex-mcp-agent@latest --yolo
```
## Tools
The MCP server exposes two tools:
- `codex_execute(prompt, work_dir)` - General purpose codex execution
- `codex_review(review_type, work_dir, target?, prompt?)` - Specialized code review
If you have any other use case requirements, feel free to open issue.
## HTTP/SSE Mode (Optional)
For local personal use, you can run the server over HTTP with Server-Sent Events (SSE) and connect via `type: "sse"`.
Start the server in SSE mode:
```bash
uvx codex-mcp-agent@latest --sse # safe mode, localhost:8822
uvx codex-mcp-agent@latest --sse --yolo # writable mode
# Options
# --host (default: 127.0.0.1)
# --port (default: 8822)
```
Configure your client `.mcp.json`:
```json
{
"mcpServers": {
"codex": {
"type": "sse",
"url": "http://127.0.0.1:8822"
}
}
}
```
Notes:
- SSE mode keeps the same Safe/Writable behavior controlled by `--yolo`.
- SSE server binds to `127.0.0.1` by default and has no auth/CORS for simplicity.
- Stdio mode remains the default; use `--sse` only if you prefer HTTP.
Optional: Streamable HTTP transport (if your client supports it):
```bash
uvx codex-mcp-agent@latest --http # safe mode streamable-http transport
```
## Safety
- **Safe Mode**: Default read-only operations protect your environment
- **Writable Mode**: Use `--yolo` flag when you need full codex capabilities
- **Sequential Execution**: Prevents conflicts from parallel agent operations
TDQS
B3.1/5.0
Scored across 3 tools
Disambiguation5/5
Each tool targets a distinct activity: general execution, repository research, and code review. No overlap in purpose.
Naming Consistency5/5
All tools follow a consistent 'codex_verb' pattern (execute, research, review), making them predictable and easy to distinguish.
Tool Count5/5
Three tools is well-scoped for a focused Codex-based agent, covering key functionalities without bloat.
Completeness4/5
The tools cover general execution, research, and code review, but could benefit from additional tools like code generation or explanation to fully round out the domain.
Maintenance
ActivityInactive
ResponsivenessNo issues