Langfuse MCP Server 🚀

🚀 Quick Start Langfuse Model Context Protocol (MCP) server enables AI agents to read/write project data directly. Choose between the native HTTP endpoint (https://langfuse.com/api/public/mcp) or the extended community CLI (npx langfuse-mcp-server).

For CLI agents (e.g., Claude Code), install the skill directly via npx skills add avivsinai/langfuse-mcp -g -y.

🔧 Configuration Bind the MCP server in your client configuration (e.g., claude_desktop_config.json). Supply LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and LANGFUSE_BASE_URL as environment variables.

{
  "mcpServers": {
    "langfuse": {
      "command": "npx",
      "args": ["langfuse-mcp-server"],
      "env": {
        "LANGFUSE_PUBLIC_KEY": "pk-...",
        "LANGFUSE_SECRET_KEY": "sk-...",
        "LANGFUSE_BASE_URL": "https://cloud.langfuse.com"
      }
    }
  }
}

📦 Dependencies Requires an MCP-compatible client like Cursor or Claude Desktop. The langfuse-mcp-server NPM package is required when executing locally via npx.

💡 Best Practices Use the MCP server for “in-the-flow” conversational debugging and live context retrieval. Restrict agent scope by allow-listing only read or list tools if write access is unnecessary.

Switch to the Langfuse CLI/SDK for heavy data processing. The SDK is superior when pre-filtering massive datasets inside a sandboxed terminal.

🚨 Gotchas / Warnings By default, the MCP server provides full read/write access to project data including scores and annotations. Secure your API keys aggressively and use environment variables rather than hardcoding them.

The native URL-based integration primarily manages prompts. You may need the community npx server for extended observability features like datasets and sessions.

🔍 Research / References

• Source: https://langfuse.com/docs/mcp
• Supported Capabilities: Observations (traces/spans), Scores (numeric/categorical logging), and Metrics (cost/latency/quality).
• Extended Objects: Datasets, dataset runs, annotation queues, comments, models, and media access.