Claude ↔ Codex MCP Relay

Claude ↔ Codex MCP Relay

A relay server that enables communication between Claude Code and Codex through shared messaging channels, facilitating collaborative workflows like code reviews. It includes tools for posting and fetching messages, a web UI for viewing conversations, and automated activity logging.

Category
访问服务器

README

Claude ↔ Codex MCP Relay (with logs + web UI)

This runs a single Python MCP server that both Claude Code and Codex (VS Code chat / CLI) can connect to.

It provides:

  • MCP tools: post_message, fetch_messages, list_channels
  • Rotating log files (tool usage + channel message archive)
  • A tiny web UI that renders conversations, highlights fenced code blocks, and can post messages

1) Setup

Requirements

  • Python 3.10+
  • pip

Install

python -m venv .venv
source .venv/bin/activate  # (Windows: .venv\Scripts\activate)
pip install -r requirements.txt

2) Run the server

python claude_codex.py

This reads host/port from config.json (or env vars, or defaults to 127.0.0.1:8010).

Alternatively, specify host/port directly:

uvicorn claude_codex:app --host 127.0.0.1 --port 8010

Open:

  • Web UI: http://127.0.0.1:8010/
  • MCP endpoint: http://127.0.0.1:8010/mcp

Logs:

  • log/claude_codex.log (rotated, defaults: 25MB × 10 backups)
  • log/channel_messages.log (rotated, defaults: 25MB × 10 backups)

3) Configuration

Configuration is loaded from config.json (if present), then environment variables, then defaults.

Option A: config.json (recommended)

Create a config.json in the project root:

{
  "host": "127.0.0.1",
  "port": 8010,
  "log_path": "log/claude_codex.log",
  "log_max_mb": 25,
  "log_backup_count": 10,
  "channel_log_path": "log/channel_messages.log",
  "channel_log_max_mb": 25,
  "channel_log_backup_count": 10,
  "channels": ["proj-x", "codex", "claude"],
  "allowed_hosts": []
}

Option B: Environment variables

Variable Default Meaning
CLAUDE_CODEX_HOST 127.0.0.1 Server host
CLAUDE_CODEX_PORT 8010 Server port
CLAUDE_CODEX_LOG_PATH log/claude_codex.log Log file path
CLAUDE_CODEX_LOG_MAX_MB 25 Rotate after N MB
CLAUDE_CODEX_LOG_BACKUP_COUNT 10 Keep N backups
CLAUDE_CODEX_CHANNEL_LOG_PATH log/channel_messages.log Channel message archive log path
CLAUDE_CODEX_CHANNEL_LOG_MAX_MB 25 Rotate channel log after N MB
CLAUDE_CODEX_CHANNEL_LOG_BACKUP_COUNT 10 Keep N channel log backups
CLAUDE_CODEX_CHANNELS proj-x,codex,claude Seed list of channels
CLAUDE_CODEX_ALLOWED_HOSTS (empty) Extra hosts for MCP DNS-rebinding protection (comma-separated)

Legacy compatibility: log_max_bytes / channel_log_max_bytes (and env ..._MAX_BYTES) are still accepted.

Example:

CLAUDE_CODEX_LOG_PATH=/tmp/relay.log uvicorn claude_codex:app --host 127.0.0.1 --port 8010

LAN / remote access

By default the MCP endpoint only accepts requests with Host: localhost or Host: 127.0.0.1 (DNS-rebinding protection). To allow connections from other machines on the network, add their IPs to allowed_hosts:

config.json:

{
  "host": "0.0.0.0",
  "allowed_hosts": ["192.168.2.3"]
}

Or via env vars:

CLAUDE_CODEX_HOST=0.0.0.0 CLAUDE_CODEX_ALLOWED_HOSTS=192.168.2.3 python claude_codex.py

Port wildcards are added automatically — 192.168.2.3 becomes 192.168.2.3:*. You can also specify an explicit host:port pattern like 192.168.2.3:8010.

4) Configure Claude Code + Codex

See:

  • docs/claude.md
  • docs/codex.md

5) Suggested workflow

Use a shared channel like proj-x so both agents read/write the same thread.

  1. Claude makes changes, then posts a review packet:

    • summary
    • unified diff (in a ```diff fenced block)
    • questions/focus
  2. Codex fetches, reviews, then posts feedback:

    • verdict
    • major/minor issues
    • tests to add/run
  3. Claude applies fixes and posts an updated packet.

6) Polling script

scripts/review-poll.sh is a lightweight bash poller that watches a channel for new messages. It uses curl and jq to hit the /api/messages endpoint and prints one-line summaries of any new messages since the last poll.

Required env vars:

Variable Purpose
BASE_URL Server URL, e.g. http://127.0.0.1:8010
LASTFILE Path to a file that persists the last-seen message ID

Optional env vars:

Variable Default Purpose
TARGET propiese Channel to poll
INTERVAL 20 Seconds between polls
WATCH_SENDER claude Only show messages from this sender (ignores own messages)

Example:

BASE_URL=http://127.0.0.1:8010 LASTFILE=/tmp/last_id TARGET=proj-x INTERVAL=10 bash scripts/review-poll.sh

7) Auto-polling: making Codex check for messages

Codex won't poll on its own — it only acts when prompted. Options A–C below require a human in the loop to relay messages to Codex. Only Option D is fully automated.

Option A: Codex custom instructions (requires human)

Add to .github/copilot-instructions.md or your VS Code Copilot instructions:

After completing any task, call the fetch_messages MCP tool on the "proj-x" channel
to check for new review feedback before going idle.

Codex will check after each task, but won't poll while idle. You still need to manually prompt Codex to start a new task cycle.

Option B: Run the poll script in a VS Code terminal (requires human)

BASE_URL=http://127.0.0.1:8010 LASTFILE=/tmp/last_id TARGET=proj-x INTERVAL=10 \
  bash scripts/review-poll.sh

New messages appear in the terminal. You need to read the output and tell Codex to act on it (e.g. "check the proj-x channel" or paste the summary).

Option C: VS Code Task (requires human)

Auto-starts the poll script when you open the workspace, so you don't have to launch it manually. But Codex still cannot read terminal output on its own — you need to prompt it when you see new messages.

Add to .vscode/tasks.json:

{
  "version": "2.0.0",
  "tasks": [
    {
      "label": "MCP Poller",
      "type": "shell",
      "command": "bash",
      "args": [".codex/reviewer-poller.sh"],
      "options": {
        "env": {
          "BASE_URL": "http://127.0.0.1:8010",
          "LASTFILE": "${workspaceFolder}/.codex/reviewer-poller.last_id",
          "TARGET": "proj-x",
          "INTERVAL": "15"
        }
      },
      "isBackground": true,
      "runOptions": { "runOn": "folderOpen" }
    }
  ]
}

Tip: Enable auto-start via Ctrl+Shift+PTasks: Manage Automatic Tasks in FolderAllow.

Option D: Codex CLI auto-loop (fully automated)

This is the only fully automated option. It uses Codex CLI — OpenAI's terminal-based coding agent — to automatically react to new messages without human intervention.

Prerequisites

  1. Install Codex CLI:

    npm i -g @openai/codex
    
  2. Configure MCP connection so Codex can call fetch_messages / post_message directly.

    Add to your project's .codex/config.toml:

    [mcp.claudecodex]
    transport = "sse"
    url = "http://127.0.0.1:8010/mcp"
    
  3. Ensure curl and jq are installed (used by the poll loop).

The auto-loop script

Create a script (e.g. .codex/auto-review-loop.sh):

The script uses codex exec (non-interactive/headless mode) so that Codex exits automatically after completing each review. The regular codex and codex resume commands open an interactive session that would block the loop.

#!/usr/bin/env bash
set -u

BASE_URL="${BASE_URL:-http://127.0.0.1:8010}"
TARGET="${TARGET:-proj-x}"
INTERVAL="${INTERVAL:-20}"
LASTFILE="${LASTFILE:-.codex/auto-loop.last_id}"
SESSION_ID="${SESSION_ID:-}"       # optional: resume an existing Codex session
WATCH_SENDER="${WATCH_SENDER:-claude}"  # only react to messages from this sender

LAST=0
if [ -f "$LASTFILE" ]; then
  LAST=$(cat "$LASTFILE" 2>/dev/null || echo 0)
fi
if ! [[ "$LAST" =~ ^[0-9]+$ ]]; then LAST=0; fi

echo "auto-loop started: target=$TARGET sender=$WATCH_SENDER interval=${INTERVAL}s last=$LAST session=${SESSION_ID:-new}"

while true; do
  JSON=$(curl -sS -m 10 "$BASE_URL/api/messages?target=$TARGET&limit=200" || true)

  # Only look at messages from WATCH_SENDER — ignore own (codex) messages
  NEW=$(printf "%s" "$JSON" | jq -r --argjson d "$LAST" --arg sender "$WATCH_SENDER" \
    '[.messages[] | select(.sender == $sender)] | last | .id // $d' 2>/dev/null || echo "$LAST")

  if [[ "$NEW" =~ ^[0-9]+$ ]] && [ "$NEW" -gt "$LAST" ]; then
    echo "$(date -Iseconds) new messages from $WATCH_SENDER (last=$LAST, new=$NEW) — launching Codex"
    PROMPT="Fetch messages from the $TARGET channel (since id $LAST) sent by $WATCH_SENDER and review them. Post your feedback back to the same channel."
    if [ -n "$SESSION_ID" ]; then
      # Resume existing session — keeps full conversation context across reviews
      codex exec --dangerously-bypass-approvals-and-sandbox resume "$SESSION_ID" "$PROMPT"
    else
      # Start a fresh session
      codex exec --dangerously-bypass-approvals-and-sandbox "$PROMPT"
    fi
    LAST="$NEW"
    echo "$LAST" > "$LASTFILE"
  fi

  sleep "$INTERVAL"
done

Run it

chmod +x .codex/auto-review-loop.sh

# Start a new session each time
BASE_URL=http://127.0.0.1:8010 TARGET=proj-x INTERVAL=15 \
  bash .codex/auto-review-loop.sh

# Or resume an existing Codex session (keeps conversation context)
BASE_URL=http://127.0.0.1:8010 TARGET=proj-x SESSION_ID=abc123 \
  bash .codex/auto-review-loop.sh

How it works

  1. The script polls /api/messages every INTERVAL seconds
  2. When a new message appears, it runs codex exec (non-interactive mode) which processes the task and exits automatically
  3. If SESSION_ID is set, it uses codex exec resume <SESSION_ID> to continue an existing session — Codex keeps the full conversation history and builds on previous reviews
  4. Codex connects to the MCP relay, fetches the new messages, reviews them, and posts feedback
  5. The last-seen ID is persisted to disk so restarts pick up where they left off
  6. The loop continues waiting for the next message

Why codex exec instead of codex? The regular codex command opens an interactive terminal UI that waits for user input and never exits. codex exec is the headless/non-interactive mode designed for scripts and automation — it runs the task and exits when done.

Codex CLI session management

Each Codex CLI chat gets a session ID stored under ~/.codex/sessions/. You can use these to maintain continuity across reviews.

Command Purpose
/status Show current session ID (inside a running session)
codex resume Pick a session to resume from a list
codex resume <SESSION_ID> Resume a specific session by ID
codex fork Fork a session into a new thread
codex fork --last Fork the most recent session

You can also type /resume inside the CLI to get an interactive session picker.

Tip: After the first Codex run, grab the session ID from /status or ~/.codex/sessions/ and pass it as SESSION_ID to the auto-loop. This way every review iteration builds on the same conversation thread, giving Codex full context of previous feedback.

Approval and sandbox modes

Flag Approvals File writes Network
(default) Asks before commands/writes Workspace only Blocked
--full-auto None Workspace only Blocked
--full-auto + network config None Workspace only Allowed
--sandbox danger-full-access Asks before commands Anywhere Allowed
--dangerously-bypass-approvals-and-sandbox None Anywhere Allowed

Important: --full-auto blocks network access by default. Since Codex needs to reach the MCP relay, the auto-loop script above will fail unless network access is enabled.

Enabling network access

Option 1: Enable network in config (recommended). Add to .codex/config.toml:

[sandbox_workspace_write]
network_access = true

Then --full-auto works as-is.

Option 2: Pass sandbox config inline:

codex exec --full-auto -c 'sandbox_workspace_write.network_access=true' "$PROMPT"

Option 3: Bypass the sandbox entirely (use with caution):

codex exec --dangerously-bypass-approvals-and-sandbox "$PROMPT"

Full-auto script with network access

This variant uses --dangerously-bypass-approvals-and-sandbox to ensure Codex can reach the MCP relay without any restrictions:

#!/usr/bin/env bash
set -u

BASE_URL="${BASE_URL:-http://127.0.0.1:8010}"
TARGET="${TARGET:-proj-x}"
INTERVAL="${INTERVAL:-20}"
LASTFILE="${LASTFILE:-.codex/auto-loop.last_id}"
SESSION_ID="${SESSION_ID:-}"
WATCH_SENDER="${WATCH_SENDER:-claude}"  # only react to messages from this sender

LAST=0
if [ -f "$LASTFILE" ]; then
  LAST=$(cat "$LASTFILE" 2>/dev/null || echo 0)
fi
if ! [[ "$LAST" =~ ^[0-9]+$ ]]; then LAST=0; fi

echo "auto-loop started: target=$TARGET sender=$WATCH_SENDER interval=${INTERVAL}s last=$LAST session=${SESSION_ID:-new}"

while true; do
  JSON=$(curl -sS -m 10 "$BASE_URL/api/messages?target=$TARGET&limit=200" || true)

  # Only look at messages from WATCH_SENDER — ignore own (codex) messages
  NEW=$(printf "%s" "$JSON" | jq -r --argjson d "$LAST" --arg sender "$WATCH_SENDER" \
    '[.messages[] | select(.sender == $sender)] | last | .id // $d' 2>/dev/null || echo "$LAST")

  if [[ "$NEW" =~ ^[0-9]+$ ]] && [ "$NEW" -gt "$LAST" ]; then
    echo "$(date -Iseconds) new messages from $WATCH_SENDER (last=$LAST, new=$NEW) — launching Codex"
    PROMPT="Fetch messages from the $TARGET channel (since id $LAST) sent by $WATCH_SENDER and review them. Post your feedback back to the same channel."
    if [ -n "$SESSION_ID" ]; then
      codex exec --dangerously-bypass-approvals-and-sandbox resume "$SESSION_ID" "$PROMPT"
    else
      codex exec --dangerously-bypass-approvals-and-sandbox "$PROMPT"
    fi
    LAST="$NEW"
    echo "$LAST" > "$LASTFILE"
  fi

  sleep "$INTERVAL"
done

Safer alternative: If you prefer to keep the workspace sandbox, add network_access = true to your .codex/config.toml (see Option 1 above) and use --full-auto instead.

8) Notes

  • This server stores messages in memory. Restarting the server clears history.
  • Want persistence? Swap the in-memory list for SQLite/Redis while keeping the same tools.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选