Skip to main content

Codex Server

The Codex MCP bridge wraps the codex exec CLI, exposing it as an MCP server with three tools: codex, code_review, and codex_reply.

Tools​

codex​

Send a prompt to OpenAI Codex for code generation or analysis.

Tool: codex
Input:
  prompt: string (required)
Output:
  string (text response)

code_review​

Send a diff for structured review. If a code reviewer agent configuration exists at ~/.codex/agents/code-reviewer.toml, its instructions are prepended to the prompt.

Tool: code_review
Input:
  diff: string (required, max 500KB)
  context: string (optional)
Output:
  JSON or text (JSON when parseable)

codex_reply​

Continue a previous conversation by passing context inline.

Tool: codex_reply
Input:
  conversation_id: string (required)
  reply: string (required)
Output:
  string (text response)
note

This tool is stateless -- context is passed inline rather than maintaining server-side conversation state.

Agent Configuration​

You can customize the code review behavior by creating a TOML file:

# ~/.codex/agents/code-reviewer.toml
[agent]
instructions = """
You are a code reviewer. Focus on:
- Security vulnerabilities
- Performance issues
- Code style consistency
"""

The path is configurable via CODEX_REVIEW_AGENT_PATH.

Configuration​

See Environment Variables for all options.

CODEX_REVIEW_MODEL=              # Model override
CODEX_REVIEW_AGENT_PATH=~/.codex/agents/code-reviewer.toml
CODEX_MCP_HTTP_PORT=8941         # HTTP proxy port