chatgpt-jules-mcp
Secure MCP server allowing ChatGPT to manage Google Jules sessions through OpenAI Secure MCP Tunnel.
README
chatgpt-jules-mcp
Secure MCP server allowing ChatGPT to manage Google Jules sessions through OpenAI Secure MCP Tunnel.
Architecture
ChatGPT
-> OpenAI Secure MCP Tunnel
-> tunnel-client (your MacBook)
-> chatgpt-jules-mcp (stdio)
-> jules-agent-sdk
-> Google Jules
-> GitHub repository
No public HTTP. No Tailscale Funnel. No raw arbitrary HTTP.
Security Features
- Policy-based access control (
policy.yaml) — allowlist sources, titles, tools - Confirmation gate for
jules.approve_plan - Audit logging (JSONL) — every tool call recorded with metadata
- Secret redaction — JULES_API_KEY, tokens, private keys scrubbed from all output
- Rate limiting — per-tool, configurable via policy
- Error boundary — all exceptions caught, sanitized, never leak secrets
- No generic HTTP tools — only Jules-specific operations
Quick Start
# Install from PyPI
pip install chatgpt-jules-mcp
# Or install in development mode
pip install -e ".[dev]"
# Set credentials
export JULES_API_KEY="your-jules-api-key"
# Configure policy
mkdir -p ~/.config/chatgpt-jules-mcp
cp config/policy.example.yaml ~/.config/chatgpt-jules-mcp/policy.yaml
# Edit: add your sources to allowed_sources
# Run MCP server (module)
python -m chatgpt_jules_mcp
# Or with CLI script (installed via pip)
chatgpt-jules-mcp
# Or with custom config
python -m chatgpt_jules_mcp --config /path/to/policy.yaml
chatgpt-jules-mcp --config /path/to/policy.yaml
MCP Tools (14)
| Tool | Category | Description |
|---|---|---|
jules.health |
Diagnostics | Server status, API key check, policy load |
jules.list_sources |
Sources | List allowed GitHub sources |
jules.get_source |
Sources | Get single source details |
jules.find_source |
Sources | Fuzzy search sources (client-side) |
jules.create_session |
Sessions | Create Jules session with policy checks |
jules.get_session |
Sessions | Get session status |
jules.list_sessions |
Sessions | List sessions (filtered by policy) |
jules.approve_plan |
Sessions | Approve plan (requires confirmation) |
jules.send_message |
Sessions | Send follow-up message |
jules.wait_for_completion |
Sessions | Poll until session completes |
jules.list_activities |
Activities | List session activities |
jules.get_activity |
Activities | Get single activity |
jules.summarize_session |
Convenience | Best-effort session summary |
jules.extract_result |
Convenience | Extract final result from activities |
Project Structure
chatgpt-jules-mcp/
├── src/chatgpt_jules_mcp/
│ ├── server.py # FastMCP server, tool registration
│ ├── policy.py # Policy engine (YAML load, validate, enforce)
│ ├── audit.py # JSONL audit logger
│ ├── redaction.py # Secret pattern redaction
│ ├── errors.py # Custom exceptions, error boundary
│ ├── ratelimit.py # In-memory rate limiter
│ ├── shutdown.py # Graceful shutdown (SIGTERM/SIGINT)
│ ├── tools/
│ │ ├── health.py # jules.health
│ │ ├── sources.py # list_sources, get_source, find_source
│ │ ├── sessions.py # create_session, get_session, list_sessions, approve_plan, send_message, wait_for_completion
│ │ ├── activities.py # list_activities, get_activity
│ │ └── convenience.py # summarize_session, extract_result
├── tests/ # 103 tests across 10 test files
├── config/
│ └── policy.example.yaml # Example policy configuration
├── scripts/
│ └── emergency-stop.sh # Kill switch
└── docs/
└── openai-secure-mcp-tunnel.md # Tunnel setup guide
Configuration
Policy is controlled via ~/.config/chatgpt-jules-mcp/policy.yaml:
jules:
allowed_sources:
- "sources/your-repo"
default_require_plan_approval: true
sessions:
allowed_title_prefixes:
- "ChatGPT:"
max_prompt_chars: 12000
max_message_chars: 8000
tools:
allow:
- "jules.health"
- "jules.list_sources"
# ... see policy.example.yaml for full list
require_confirmation:
- "jules.approve_plan"
audit:
path: "~/.local/share/chatgpt-jules-mcp/audit.jsonl"
redact_secrets: true
Tests
pip install pytest pytest-asyncio
PYTHONPATH=src python -m pytest tests/ -v
50 tests covering: policy engine, audit logging, redaction, rate limiting, all 14 tools, error handling, input validation, secret redaction in transit.
OpenAI Secure MCP Tunnel
See docs/openai-secure-mcp-tunnel.md for full setup.
Quick version:
export CONTROL_PLANE_API_KEY="sk-..."
export CONTROL_PLANE_TUNNEL_ID="tunnel_..."
export JULES_API_KEY="..."
tunnel-client init \
--sample sample_mcp_stdio_local \
--profile chatgpt-jules-mcp-local \
--tunnel-id "$CONTROL_PLANE_TUNNEL_ID" \
--mcp-command "python -m chatgpt_jules_mcp"
tunnel-client doctor --profile chatgpt-jules-mcp-local --explain
tunnel-client run --profile chatgpt-jules-mcp-local
Emergency Stop
./scripts/emergency-stop.sh
Production Deployment
Security Checklist
- [ ] Set
allowed_sourcesto specific repositories (nevernullin production) - [ ] Set
allowed_title_prefixesto restrict session titles - [ ] Disable
log_full_prompts(keepfalse) - [ ] Set
redact_secrets: true - [ ] Restrict audit log file permissions (
chmod 600 audit.jsonl) - [ ] Run behind OpenAI Secure MCP Tunnel (no public HTTP)
- [ ] Set rate limits appropriate for your usage (
calls_per_minute) - [ ] Review and customize
redaction.patternsfor your environment
systemd Service (Linux)
Create /etc/systemd/system/chatgpt-jules-mcp.service:
[Unit]
Description=ChatGPT Jules MCP Server
After=network.target
[Service]
Type=simple
User=chatgpt-jules
Group=chatgpt-jules
Environment="JULES_API_KEY=your-key-here"
Environment="PYTHONUNBUFFERED=1"
WorkingDirectory=/opt/chatgpt-jules-mcp
ExecStart=/opt/chatgpt-jules-mcp/.venv/bin/python -m chatgpt_jules_mcp
Restart=on-failure
RestartSec=5
[Install]
WantedBy=multi-user.target
Enable and start:
sudo systemctl daemon-reload
sudo systemctl enable chatgpt-jules-mcp
sudo systemctl start chatgpt-jules-mcp
sudo systemctl status chatgpt-jules-mcp
Docker
FROM python:3.11-slim
WORKDIR /app
COPY . .
RUN pip install -e ".[dev]"
ENV PYTHONUNBUFFERED=1
USER 1000
CMD ["python", "-m", "chatgpt_jules_mcp"]
docker build -t chatgpt-jules-mcp .
docker run -e JULES_API_KEY=... chatgpt-jules-mcp
License
Apache-2.0
推荐服务器
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。
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。