langobs-mcp

# langobs-mcp MCP server for [LangObs](https://github.com/langobs/langobs) — zero-code observability for any MCP-compatible agent (Claude Desktop, Cursor, Windsurf, etc.). Any agent that supports MCP can send traces to LangObs without installing the Python or TypeScript SDK. ## Install ```bash pip install langobs-mcp ``` ## Prerequisites Start the LangObs server: ```bash pip install langobs-server langobs-server ``` ## Configure in your MCP client ### Claude Desktop (`~/Library/Application Support/Claude/claude_desktop_config.json`) ```json { "mcpServers": { "langobs": { "command": "langobs-mcp" } } } ``` ### Cursor / Windsurf Add to your MCP server config: ```json { "langobs": { "command": "langobs-mcp" } } ``` ## Tools Exposed Once connected, your MCP agent can call these tools: | Tool | Description | |------|-------------| | `langobs_start_session` | Start a new debugging session. Returns a `session_id` to use in subsequent calls. | | `langobs_report_trace` | Report a trace event (LLM call, tool call, decision, memory op, error). | | `langobs_report_memory` | Report a memory state change (created, updated, accessed, deleted). | ## How It Works The MCP server receives tool calls from your agent and forwards them to the LangObs HTTP server (`http://localhost:8766`). Events appear in the dashboard in real-time. ``` Agent → MCP call → langobs-mcp server → HTTP POST → LangObs server → Dashboard ``` ## Example Usage (inside a Claude conversation) Once the MCP server is connected, Claude can instrument its own tasks: ``` 1. Call langobs_start_session → get session_id 2. For each tool call: langobs_report_trace(session_id, event_type="tool_call", ...) 3. For each LLM decision: langobs_report_trace(session_id, event_type="decision", ...) ``` Open `http://localhost:5173` to watch the trace build in real time. ## License MIT