ToolTrace
zero-config call tracing for your MCP tool servers
npx -y github:Isco0819/tool-traceEver run into this?
Like Langfuse which has real traction ($2M+ ARR, Product Hunt launch) for LLM observability, but no equivalent exists at the MCP tool-call level. When your Claude agent fires 40+ tool calls per session you get zero visibility into which tools are slow, which fail silently, or which get called redundantly — debugging agent harnesses is pure guesswork.
The fix
A drop-in npx MCP proxy that transparently wraps any stdio MCP server with zero config changes, logs structured JSONL traces locally, and renders a live timing table in terminal so you can see exactly what your agent is doing.
What it does
- Transparent stdio proxy: sits between Claude and any MCP server, zero changes to existing server code
- Per-call JSONL traces to ~/.tooltrace/: tool name, args hash, duration ms, success/error, session ID
- Live ink.js terminal dashboard: slowest tools ranked, error rate per tool, duplicate-call detector