SnapAgent

Snapshot-test your AI agent's tool-call behavior and catch silent regressions

npx -y github:Isco0819/snapagent

Ever run into this?

Like jest-snapshot (part of Jest's 43k-star ecosystem) which locks down expected UI output, but nothing equivalent exists for AI agent behavior. When you upgrade an MCP server, tweak a system prompt, or bump your agent framework, you have zero visibility into whether the agent started calling different tools, in a different order, or with different args. Regressions in agent behavior ship silently.

The fix

npx snapagent wraps your MCP stdio server as a transparent proxy, recording every tool call and response schema into a JSON snapshot file. On the next run it diffs the two sessions and exits 1 on drift — drop it into CI like any snapshot test.

What it does

npx -y github:Isco0819/snapagent