Hey Product Hunt! I'm the maker of Heimdall MCP. I built this after hitting the same wall every time I ran AI agents with MCP tools: something was slow, or failing silently, and I had zero visibility into what actually happened inside the server. Which tool did Claude call? What input did it send? How long did it take? No way to know. Heimdall sits between your MCP client (Claude, OpenCode, Cursor) and any MCP server — local or remote — and records every tool call as an OpenTelemetry span. No modifications to the target server. Just wrap it in mcp.json with one line change: "command": "heimdall", "args": ["--store", "sqlite://~/.heimdall/traces.db", "--", "node", "my-server.js"] Storage options: SQLite (local, zero infra), Postgres, MySQL, or any OTLP-compatible backend. Pure JS/WASM — no native deps, no node-gyp. Also ships as a TypeScript library with a fluent builder API and custom interceptors if you want tighter integration. Happy to answer questions about the OTel schema design or the transport mixing approach (stdio in → HTTP out). Those were the trickiest parts.

Heimdall MCP

Trace every MCP tool call. No code changes needed.

Trace every MCP tool call. No code changes needed.