Gemini MCP Server

# Getting Started with Zen MCP Server This guide walks you through setting up the Zen MCP Server from scratch, including installation, configuration, and first usage. ## Prerequisites - **Python 3.10+** (3.12 recommended) - **Git** - **[uv installed](https://docs.astral.sh/uv/getting-started/installation/)** (for uvx method) - **Windows users**: WSL2 required for Claude Code CLI ## Step 1: Get API Keys You need at least one API key. Choose based on your needs: ### Option A: OpenRouter (Recommended for beginners) **One API for multiple models** - Visit [OpenRouter](https://openrouter.ai/) and sign up - Generate an API key - Control spending limits in your dashboard - Access GPT-4, Claude, Gemini, and more through one API ### Option B: Native Provider APIs **Gemini (Google):** - Visit [Google AI Studio](https://makersuite.google.com/app/apikey) - Generate an API key - **Note**: For Gemini 2.5 Pro, use a paid API key (free tier has limited access) **OpenAI:** - Visit [OpenAI Platform](https://platform.openai.com/api-keys) - Generate an API key for O3, GPT-5 access **X.AI (Grok):** - Visit [X.AI Console](https://console.x.ai/) - Generate an API key for Grok models **DIAL Platform:** - Visit [DIAL Platform](https://dialx.ai/) - Generate API key for vendor-agnostic model access ### Option C: Local Models (Free) **Ollama:** ```bash # Install Ollama curl -fsSL https://ollama.ai/install.sh | sh # Start Ollama service ollama serve # Pull a model (e.g., Llama 3.2) ollama pull llama3.2 ``` **Other local options:** - **vLLM**: Self-hosted inference server - **LM Studio**: Local model hosting with OpenAI-compatible API - **Text Generation WebUI**: Popular local interface 👉 **[Complete custom model setup guide](custom_models.md)** ## Step 2: Installation Choose your preferred installation method: ### Method A: Instant Setup with uvx (Recommended) **Prerequisites**: [Install uv first](https://docs.astral.sh/uv/getting-started/installation/) Choose your AI coding assistant and add the corresponding configuration: **For Claude Desktop:** 1. Open Claude Desktop → Settings → Developer → Edit Config 2. Add this configuration: ```json { "mcpServers": { "zen": { "command": "sh", "args": [ "-c", "for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x \"$p\" ] && exec \"$p\" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1" ], "env": { "PATH": "/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:~/.local/bin", "GEMINI_API_KEY": "your_api_key_here" } } } } ``` **For Claude Code CLI:** Create `.mcp.json` in your project root: ```json { "mcpServers": { "zen": { "command": "sh", "args": [ "-c", "for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x \"$p\" ] && exec \"$p\" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1" ], "env": { "PATH": "/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:~/.local/bin", "GEMINI_API_KEY": "your_api_key_here" } } } } ``` **For Gemini CLI:** Edit `~/.gemini/settings.json`: ```json { "mcpServers": { "zen": { "command": "sh", "args": [ "-c", "for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x \"$p\" ] && exec \"$p\" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1" ], "env": { "PATH": "/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:~/.local/bin", "GEMINI_API_KEY": "your_api_key_here" } } } } ``` **For Codex CLI:** Edit `~/.codex/config.toml`: ```toml [mcp_servers.zen] command = "bash" args = ["-c", "for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x \\\"$p\\\" ] && exec \\\"$p\\\" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1"] [mcp_servers.zen.env] PATH = "/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:$HOME/.local/bin:$HOME/.cargo/bin:$HOME/bin" GEMINI_API_KEY = "your_api_key_here" ``` Enable Codex's built-in web-search tool so Zen's `apilookup` instructions can execute successfully: ```toml [tools] web_search = true ``` Add the block above if `[tools]` is missing from the file; otherwise ensure `web_search = true` appears in that section. **For Qwen Code CLI:** Create or edit `~/.qwen/settings.json`: ```json { "mcpServers": { "zen": { "command": "bash", "args": [ "-c", "for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x \"$p\" ] && exec \"$p\" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1" ], "cwd": "/path/to/zen-mcp-server", "env": { "PATH": "/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:~/.local/bin", "GEMINI_API_KEY": "your_api_key_here" } } } } ``` Replace the placeholder API key with the providers you use (Gemini, OpenAI, OpenRouter, etc.). **For OpenCode CLI:** Edit `~/.config/opencode/opencode.json`: ```json { "$schema": "https://opencode.ai/config.json", "mcp": { "zen": { "type": "local", "command": [ "/path/to/zen-mcp-server/.zen_venv/bin/python", "/path/to/zen-mcp-server/server.py" ], "cwd": "/path/to/zen-mcp-server", "enabled": true, "environment": { "GEMINI_API_KEY": "your_api_key_here" } } } } ``` Add any other API keys you rely on (`OPENAI_API_KEY`, `OPENROUTER_API_KEY`, etc.). #### IDE Clients (Cursor & VS Code) Zen works in GUI IDEs that speak MCP. The configuration mirrors the CLI examples above—point the client at the `uvx` launcher and set any required environment variables. **Cursor IDE** 1. Open Cursor → `Settings` (`Cmd+,`/`Ctrl+,`) → **Integrations › Model Context Protocol (MCP)**. 2. Click **Add MCP Server** and supply the following values: - Command: `sh` - Args: `-c` and `for p in $(which uvx 2>/dev/null) $HOME/.local/bin/uvx /opt/homebrew/bin/uvx /usr/local/bin/uvx uvx; do [ -x "$p" ] && exec "$p" --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server; done; echo 'uvx not found' >&2; exit 1` - Environment (example): - `PATH=/usr/local/bin:/usr/bin:/bin:/opt/homebrew/bin:~/.local/bin` - `GEMINI_API_KEY=your_api_key_here` 3. Save the configuration—Cursor will launch the MCP server on demand. See the [Cursor MCP guide](https://cursor.com/docs) for screenshots of the UI. **Visual Studio Code (Claude Dev extension)** 1. Install the [Claude Dev extension](https://marketplace.visualstudio.com/items?itemName=Anthropic.claude-vscode) v0.6.0 or later. 2. Open the Command Palette (`Cmd+Shift+P`/`Ctrl+Shift+P`) → **Claude: Configure MCP Servers** → **Add server**. 3. When prompted, use the same values as above: - Command: `sh` - Args: `-c` and the `uvx` bootstrap loop - Environment: add the API keys you need (e.g. `GEMINI_API_KEY`, `OPENAI_API_KEY`) 4. Save the JSON snippet the extension generates. VS Code will reload the server automatically the next time you interact with Claude. 👉 Pro tip: If you prefer a one-line command, replace the long loop with `uvx --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server`—just make sure `uvx` is on your PATH for every client. **Benefits of uvx method:** - ✅ Zero manual setup required - ✅ Always pulls latest version - ✅ No local dependencies to manage - ✅ Works without Python environment setup ### Method B: Clone and Setup ```bash # Clone the repository git clone https://github.com/BeehiveInnovations/zen-mcp-server.git cd zen-mcp-server # One-command setup (handles everything) ./run-server.sh # Or for Windows PowerShell: ./run-server.ps1 # View configuration for Claude Desktop ./run-server.sh -c # See all options ./run-server.sh --help ``` **What the setup script does:** - ✅ Creates Python virtual environment - ✅ Installs all dependencies - ✅ Creates .env file for API keys - ✅ Configures Claude integrations - ✅ Provides copy-paste configuration **After updates:** Always run `./run-server.sh` again after `git pull`. **Windows users**: See the [WSL Setup Guide](wsl-setup.md) for detailed WSL configuration. ## Step 3: Configure API Keys ### For uvx installation: Add your API keys directly to the MCP configuration shown above. ### For clone installation: Edit the `.env` file: ```bash nano .env ``` Add your API keys (at least one required): ```env # Choose your providers (at least one required) GEMINI_API_KEY=your-gemini-api-key-here # For Gemini models OPENAI_API_KEY=your-openai-api-key-here # For O3, GPT-5 XAI_API_KEY=your-xai-api-key-here # For Grok models OPENROUTER_API_KEY=your-openrouter-key # For multiple models # DIAL Platform (optional) DIAL_API_KEY=your-dial-api-key-here DIAL_API_HOST=https://core.dialx.ai # Default host (optional) DIAL_API_VERSION=2024-12-01-preview # API version (optional) DIAL_ALLOWED_MODELS=o3,gemini-2.5-pro # Restrict models (optional) # Custom/Local models (Ollama, vLLM, etc.) CUSTOM_API_URL=http://localhost:11434/v1 # Ollama example CUSTOM_API_KEY= # Empty for Ollama CUSTOM_MODEL_NAME=llama3.2 # Default model name ``` ## Prevent Client Timeouts Some MCP clients default to short timeouts and can disconnect from Zen during long tool runs. Configure each client to wait at least five minutes (300 000 ms) before giving up. ### Claude Code & Claude Desktop Claude reads MCP-related environment variables either from your shell or from `~/.claude/settings.json`. Add (or update) the `env` block so both startup and tool execution use a 5-minute limit: ```json { "env": { "MCP_TIMEOUT": "300000", "MCP_TOOL_TIMEOUT": "300000" } } ``` You can scope this block at the top level of `settings.json` (applies to every session) or under a specific `mcpServers.<name>.env` entry if you only want it for Zen. The values are in milliseconds. Note: Claude’s SSE transport still enforces an internal ceiling of roughly five minutes; long-running HTTP/SSE servers may need retries until Anthropic ships their fix. ### Codex CLI Codex exposes per-server timeouts in `~/.codex/config.toml`. Add (or bump) these keys under `[[mcp_servers.<name>]]`: ```toml [mcp_servers.zen] command = "..." args = ["..."] startup_timeout_sec = 300 # default is 10 seconds tool_timeout_sec = 300 # default is 60 seconds ``` `startup_timeout_sec` covers the initial handshake/list tools step, while `tool_timeout_sec` governs each tool call. Increase either beyond 300 if your workflow routinely exceeds five minutes. ### Gemini CLI Gemini uses a single `timeout` field per server inside `~/.gemini/settings.json`. Set it to at least five minutes (values are milliseconds): ```json { "mcpServers": { "zen": { "command": "uvx", "args": ["zen-mcp-server"], "timeout": 300000 } } } ``` Versions 0.2.1 and newer currently ignore values above ~60 seconds for some transports due to a known regression; if you still see premature disconnects we recommend breaking work into smaller calls or watching the Gemini CLI release notes for the fix. **Important notes:** - ⭐ **No restart needed** - Changes take effect immediately - ⭐ If multiple APIs configured, native APIs take priority over OpenRouter - ⭐ Configure model aliases in [`conf/custom_models.json`](../conf/custom_models.json) ## Step 4: Test the Installation ### For Claude Desktop: 1. Restart Claude Desktop 2. Open a new conversation 3. Try: `"Use zen to list available models"` ### For Claude Code CLI: 1. Exit any existing Claude session 2. Run `claude` from your project directory 3. Try: `"Use zen to chat about Python best practices"` ### For Gemini CLI: **Note**: While Zen MCP connects to Gemini CLI, tool invocation isn't working correctly yet. See [Gemini CLI Setup](gemini-setup.md) for updates. ### For Qwen Code CLI: 1. Restart the Qwen Code CLI if it's running (`qwen exit`). 2. Run `qwen mcp list --scope user` and confirm `zen` shows `CONNECTED`. 3. Try: `"/mcp"` to inspect available tools or `"Use zen to analyze this repo"`. ### For OpenCode CLI: 1. Restart OpenCode (or run `OpenCode: Reload Config`). 2. Open **Settings › Tools › MCP** and confirm `zen` is enabled. 3. Start a new chat and try: `"Use zen to list available models"`. ### For Codex CLI: 1. Restart Codex CLI if running 2. Open a new conversation 3. Try: `"Use zen to list available models"` ### Test Commands: ``` "Use zen to list available models" "Chat with zen about the best approach for API design" "Use zen thinkdeep with gemini pro about scaling strategies" "Debug this error with o3: [paste error]" ``` **Note**: Codex CLI provides excellent MCP integration with automatic environment variable configuration when using the setup script. ## Step 5: Start Using Zen ### Basic Usage Patterns: **Let Claude pick the model:** ``` "Use zen to analyze this code for security issues" "Debug this race condition with zen" "Plan the database migration with zen" ``` **Specify the model:** ``` "Use zen with gemini pro to review this complex algorithm" "Debug with o3 using zen for logical analysis" "Get flash to quickly format this code via zen" ``` **Multi-model workflows:** ``` "Use zen to get consensus from pro and o3 on this architecture" "Code review with gemini, then precommit validation with o3" "Analyze with flash, then deep dive with pro if issues found" ``` ### Quick Tool Reference: **🤝 Collaboration**: `chat`, `thinkdeep`, `planner`, `consensus` **🔍 Code Analysis**: `analyze`, `codereview`, `debug`, `precommit` **⚒️ Development**: `refactor`, `testgen`, `secaudit`, `docgen` **🔧 Utilities**: `challenge`, `tracer`, `listmodels`, `version` 👉 **[Complete Tools Reference](tools/)** with detailed examples and parameters ## Common Issues and Solutions ### "zen not found" or "command not found" **For uvx installations:** - Ensure `uv` is installed and in PATH - Try: `which uvx` to verify uvx is available - Check PATH includes `/usr/local/bin` and `~/.local/bin` **For clone installations:** - Run `./run-server.sh` again to verify setup - Check virtual environment: `which python` should show `.zen_venv/bin/python` ### API Key Issues **"Invalid API key" errors:** - Verify API keys in `.env` file or MCP configuration - Test API keys directly with provider's API - Check for extra spaces or quotes around keys **"Model not available":** - Run `"Use zen to list available models"` to see what's configured - Check model restrictions in environment variables - Verify API key has access to requested models ### Performance Issues **Slow responses:** - Use faster models: `flash` instead of `pro` - Lower thinking modes: `minimal` or `low` instead of `high` - Restrict model access to prevent expensive model selection **Token limit errors:** - Use models with larger context windows - Break large requests into smaller chunks - See [Working with Large Prompts](advanced-usage.md#working-with-large-prompts) ### More Help 👉 **[Complete Troubleshooting Guide](troubleshooting.md)** with detailed solutions 👉 **[Advanced Usage Guide](advanced-usage.md)** for power-user features 👉 **[Configuration Reference](configuration.md)** for all options ## What's Next? 🎯 **Try the example workflows in the main README** 📚 **Explore the [Tools Reference](tools/)** to understand what each tool can do ⚡ **Read the [Advanced Usage Guide](advanced-usage.md)** for complex workflows 🔧 **Check out [Configuration Options](configuration.md)** to customize behavior 💡 **Join discussions and get help** in the project issues or discussions ## Quick Configuration Templates ### Development Setup (Balanced) ```env DEFAULT_MODEL=auto GEMINI_API_KEY=your-key OPENAI_API_KEY=your-key GOOGLE_ALLOWED_MODELS=flash,pro OPENAI_ALLOWED_MODELS=o4-mini,o3-mini ``` ### Cost-Optimized Setup ```env DEFAULT_MODEL=flash GEMINI_API_KEY=your-key GOOGLE_ALLOWED_MODELS=flash ``` ### High-Performance Setup ```env DEFAULT_MODEL=auto GEMINI_API_KEY=your-key OPENAI_API_KEY=your-key GOOGLE_ALLOWED_MODELS=pro OPENAI_ALLOWED_MODELS=o3 ``` ### Local-First Setup ```env DEFAULT_MODEL=auto CUSTOM_API_URL=http://localhost:11434/v1 CUSTOM_MODEL_NAME=llama3.2 # Add cloud APIs as backup GEMINI_API_KEY=your-key ``` Happy coding with your AI development team! 🤖✨