mcp-openai

MCP server wrapping OpenAI Codex CLI for non-interactive API calls with error detection.

Tools

| Tool | Description | |------|-------------| | openai_chat | Send prompts via codex exec (fast, non-interactive) | | openai_review | Code review via codex review (git-based) |

Prerequisites

1. Install Codex CLI:

   npm install -g @openai/codex

2. Authenticate:

   codex login

3. Create minimal config (for fast startup without MCP servers):

   mkdir -p ~/.codex-minimal
   cp ../../config/codex-minimal.toml.example ~/.codex-minimal/config.toml
   # Link auth from main codex config:
   ln -s ~/.codex/auth.json ~/.codex-minimal/auth.json

Configuration

The server uses CODEX_HOME env var to find the Codex config directory. Defaults to ~/.codex-minimal.

Set it in your .mcp.json:

{
  "mcp-openai": {
    "type": "stdio",
    "command": "node",
    "args": ["/path/to/servers/mcp-openai/server.js"],
    "env": {
      "CODEX_HOME": "/path/to/.codex-minimal"
    }
  }
}

Error Detection

The server automatically detects and reports:

• QUOTA_EXCEEDED — ChatGPT Plus weekly credit limit reached
• MODEL_NOT_SUPPORTED — model not available on your plan
• AUTH_EXPIRED — OAuth token needs refresh

Why Minimal Config?

The default codex mcp-server loads ALL configured MCP servers on startup, making it slow and unreliable. This server uses codex exec with a separate CODEX_HOME that has no MCP servers configured, resulting in instant startup.