ReplayMCP
Record & replay MCP tool calls for deterministic agent testing
npx -y github:Isco0819/replay-mcpEver run into this?
Like vcr.py (6k GitHub stars) and nock (13k stars) which proved massive demand for HTTP record/replay, no equivalent exists for MCP tool calls. Every CI run of your agent harness hits live tools — slow, expensive, and non-deterministic. Flaky agent tests are the number-one blocker when shipping AI-native apps.
The fix
A transparent MCP stdio proxy that records every tool call request/response pair to a JSON cassette file in record mode, then serves those snapshots in replay mode — zero real API calls needed after the first run.
What it does
- record mode: wrap any stdio MCP server and snapshot every tool call I/O to cassette.json keyed by tool-name+params-hash
- replay mode: serve cassette responses fully offline — deterministic, instant, free
- diff mode: run live vs cassette side-by-side and surface API-shape drift before it silently breaks your agent