@popcornai/mcp

An MCP server for Popcorn AI. It lets Claude, or any MCP client, search your AI agent's runs, read a trace step by step, and run an eval suite.

Add to Claude Code

claude mcp add popcorn \
  --env POPCORN_ENDPOINT=http://localhost:4100 \
  --env POPCORN_API_KEY=pk_... \
  -- npx -y @popcornai/mcp

Other clients: run npx -y @popcornai/mcp as a stdio server with the same two environment variables.

Tools

| Tool | What it does | | ------------- | ------------------------------------------------------------------------------------------------- | | search_runs | Find runs by text, status, flag (error, loop, cost_spike, max_steps) or agent name | | get_run | A run step by step: each model and tool call with arguments, results and errors | | run_eval | Run a suite against the agent in popcorn.config.ts and return the review. Calls your real model |

Payloads are clipped to 1,500 characters each so long runs fit in context.

Try: "Find the runs flagged as loops this week and tell me what the agent kept retrying."

Requires Node 20.11 or newer and a running Popcorn server. MIT licensed.