claude-code-codex-agents

日本語版 README はこちら

Give Claude Code structured Codex traces, not raw output.

For Claude Code users who want GPT-5.4 as a real tool: claude-code-codex-agents parses the entire JSONL event stream from Codex CLI and returns a structured execution report -- which tools it used, which files it touched, how long it took, and what went wrong. No other Codex MCP bridge does this.

graph LR
    A["Claude Code<br/>(Opus 4.6)"] -->|MCP Protocol| B["claude-code-codex-agents<br/>MCP Server"]
    B -->|"subprocess + stdin"| C[Codex CLI]
    C -->|JSONL stream| B
    C -->|API call| D["OpenAI API<br/>(GPT-5.4)"]
    B -->|Structured Report| A

Without vs With claude-code-codex-agents

Without -- You call Codex CLI and get a wall of text. You don't know what tools it used, what files it changed, or if it actually succeeded.

With claude-code-codex-agents -- Claude Code gets a structured execution trace:

[Codex gpt-5.4] Completed

⏱ Execution time: 8.3s
🧵 Thread: 019d436e-4c39-7093-b7ed-f8a26aca7938

📦 Tools used (3):
  ✅ read_file — src/auth.py
  ✅ edit_file — src/auth.py
  ✅ shell — python -m pytest tests/

📁 Files touched (1):
  • src/auth.py

━━━ Codex Response ━━━
Fixed the authentication logic. Token validation order was incorrect.

Related MCP server: Codex MCP Server

Why claude-code-codex-agents?

There are 6+ Codex MCP bridges on GitHub. Here's what makes this one different:

Key Features

• Full JSONL Trace Parsing -- Every Codex event (tool calls, file ops, errors) parsed into a structured report

• Parallel Execution -- Run up to 6 Codex tasks simultaneously via parallel_execute

• Session Management -- Continue previous threads with session_continue (threadId persistence)

• Agent Lifecycle -- Run Codex as a background Claude Code-style worker via spawn_codex_agent, send_codex_agent_input, and wait_codex_agent

• Adversarial Review Loop -- GPT-5.4 reviews Claude's code from a different perspective

• Sandbox Security -- 3-tier policy (read-only / workspace-write / danger-full-access) + terminal injection prevention

• Cross-Model Discussion -- Get GPT-5.4's opinion on design decisions via discuss

• Zero External Dependencies -- Just FastMCP + Codex CLI. No databases, no Docker, no config files

• Japanese Native -- Full Japanese prompt and report support

• 59 Tests -- Comprehensive coverage including security, parsing, session management, agent lifecycle, and edge cases

Quick Start

1. Install Codex CLI

npm install -g @openai/codex
codex login

2. Install claude-code-codex-agents

git clone https://github.com/tsunamayo7/claude-code-codex-agents.git
cd claude-code-codex-agents
uv sync

3. Add to your MCP client

Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "claude-code-codex-agents": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "--directory", "/path/to/claude-code-codex-agents", "python", "server.py"],
      "env": { "PYTHONUTF8": "1" }
    }
  }
}

{
  "mcpServers": {
    "claude-code-codex-agents": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/claude-code-codex-agents", "python", "server.py"],
      "env": { "PYTHONUTF8": "1" }
    }
  }
}

Add to your MCP settings:

{
  "claude-code-codex-agents": {
    "command": "uv",
    "args": ["run", "--directory", "/path/to/claude-code-codex-agents", "python", "server.py"],
    "env": { "PYTHONUTF8": "1" }
  }
}

Tools

Claude Code-Style Agents

The new agent lifecycle tools let Claude Code treat Codex more like a persistent sub-agent than a one-shot CLI call.

• Use spawn_codex_agent to start a background worker with a role preset: default for balanced execution, explorer for read-heavy investigation, worker for implementation.

• Use send_codex_agent_input to continue the same worker after you read its last result.

• Use wait_codex_agent to poll for completion without blocking other work.

• Use list_codex_agents and close_codex_agent to manage idle workers.

Real-World Example: Adversarial Code Review

Claude Code writes code, then asks GPT-5.4 to review it:

[Codex Review] GPT-5.4 Review Result

⏱ Execution time: 15.7s

━━━ Codex Response ━━━
- [CRITICAL] `run(cmd)` calls `os.system(cmd)` directly -- command injection
  if `cmd` contains user input. Use `subprocess.run([...], shell=False)`.

- [WARNING] `divide(a, b)` raises ZeroDivisionError when b == 0.
  Add a pre-check or explicit error message.

- [INFO] No type hints on function signatures. Add `def divide(a: float,
  b: float) -> float:` for readability.

Real-World Example: Parallel Execution

Analyze multiple tasks simultaneously:

[Parallel Execution Complete] 3 tasks

━━━ Task 1 ✅ ━━━
Instruction: Analyze src/auth.py for security issues
⏱ 5.2s
...

━━━ Task 2 ✅ ━━━
Instruction: Review database query patterns in src/db.py
⏱ 7.8s
...

━━━ Task 3 ✅ ━━━
Instruction: Check error handling in src/api.py
⏱ 4.1s
...

Architecture

sequenceDiagram
    participant C as Claude Code
    participant H as claude-code-codex-agents
    participant X as Codex CLI
    participant O as OpenAI API

    C->>H: MCP tool call (execute)
    H->>H: _validate() + _enforce_sandbox()
    H->>X: subprocess (stdin prompt)
    X->>O: API request (GPT-5.4)
    O-->>X: Response
    X-->>H: JSONL event stream
    H->>H: parse_jsonl_events() → CodexTrace
    H->>H: _sanitize() → format_report()
    H-->>C: Structured report

Security Model

Additional protections:

• ANSI/OSC escape sequence sanitization (terminal injection prevention)

• Input validation on all parameters

• Process kill on timeout

• --ephemeral flag (no persistent Codex state)

Development

# Setup
git clone https://github.com/tsunamayo7/claude-code-codex-agents.git
cd claude-code-codex-agents
uv sync --extra dev

# Run tests (59 tests)
uv run pytest tests/ -v

# Run server directly
uv run python server.py

Project structure: Single file (server.py, ~820 lines). Easy to read, modify, and contribute.

Use Cases

1. Cross-Model Code Review -- Claude writes code, GPT-5.4 reviews it. Eliminates single-model bias.

2. Parallel Codebase Analysis -- Analyze 6 files simultaneously, get structured reports for each.

3. Design Discussion -- Get GPT-5.4's alternative perspective on architectural decisions via discuss.

4. Session-Based Refactoring -- Large refactoring across multiple session_continue calls with context preservation.

5. AI Second Opinion -- When Claude's answer seems off, ask GPT-5.4 for a sanity check.

Requirements

• Python 3.12+

• Codex CLI (npm install -g @openai/codex)

• OpenAI account (Codex CLI must be authenticated via codex login)

• uv (recommended) or pip

Related Projects

Helix Ecosystem

• helix-ai-studio — All-in-one AI chat studio with 7 providers, RAG, MCP tools, and pipeline

• helix-pilot — GUI automation MCP server — AI controls Windows desktop via local Vision LLM

• helix-agent — Extend Claude Code with local Ollama models — cut token costs by 60-80%

• helix-sandbox — Secure sandbox MCP server — Docker + Windows Sandbox

Alternative Codex Bridges

• codex-plugin-cc -- Official OpenAI plugin for Claude Code

• codex-mcp-server -- Alternative Codex MCP bridge (Node.js)

License

MIT