When an AI agent verifies something ("POST /orders works, the row is there"), that knowledge usually dies with the chat. Next sprint, someone asks again.
The fix is boring and effective: make the agent save what it verified, as files in the repo.
The loop
With the SignOff Studio MCP server connected (Claude Code, Cursor, VS Code Copilot or Gemini CLI):
- Verify.
verify_api_effectruns a request plus database checks bound to its response, and returns one verdict. - Save.
save_requestandsave_db_checkwrite the request, its assertions and the database check into the workspace (signoff-workspace.json, collections,database/). These are plain files. - Commit. They go in the same pull request as the code.
- Rerun.
run_collection(from the agent), the desktop app, the VS Code extension or the CLI in CI runs them again. Failures show expected vs actual. - Report.
export_evidencewrites an HTML/Markdown report plus JUnit, and a short summary to paste in the PR.
Prompts that work
The server ships slash-command prompts in Claude Code and VS Code:
-
verify-endpoint: call it, check it, save it. -
verify-migration: check the schema and data after a migration. -
regression-check: rerun the saved collection and summarise failures. -
explore-api: import an OpenAPI/Postman/Insomnia/Bruno/cURL file and try the main endpoints.
Why files, not a SaaS dashboard
- Reviewers see the checks in the diff.
- CI runs them with no account and no network dependency.
- Nothing leaves your machine: requests and queries run locally, and results go only to your agent.
Getting started
npx -y @signoff-studio/mcp init Node 22+, or use the free Windows app / VS Code extension, which ship the same server (Help → Connect AI agents writes the config, with a preview of every file).
testapilocally.com/mcp · GitHub
Disclosure: I build SignOff Studio.