Letting Claude Code test your API locally
How do I let Claude Code test my API locally?
Add an MCP server that makes HTTP requests and checks the results, then ask Claude Code to use it. With SignOff Studio:
claude mcp add signoff-studio -- npx -y @signoff-studio/mcp
(On native Windows: -- cmd /c npx -y @signoff-studio/mcp.) Then ask: "Use signoff-studio to check POST /orders returns 201 and saves the order." Claude Code calls http_request with checks, or verify_api_effect to check the database row too, and gets a short verdict back.
Why not just let it use curl?
It can, and for a single quick look curl is fine. The trouble shows up over a whole task:
- Claude has to read raw output, including headers and big JSON, which eats its context.
- There's no shared notion of "pass": every check is a new shell pipeline.
- Tokens and passwords end up in commands, shell history and the transcript.
- Nothing is saved. The next session, or your teammate, starts over.
The MCP tools return a compact result with pass or fail per check, redact secrets by value, and can save the check (save_request) into a workspace that's committed with the code. More in curl vs MCP for API testing.
Does my API need to be public?
No. The server runs on your machine and calls localhost, private network addresses and anything else your machine can reach. Nothing goes through a cloud relay.
How does Claude Code know my base URL and tokens?
If the project has a SignOff Studio workspace, Claude uses its environments ({{baseUrl}}, {{token}}). Secret values come from the workspace .env, so they never appear in committed files or in the model's context. Without a workspace, Claude passes full URLs and headers itself.
Can it run my existing Postman collection?
Yes. import_requests imports Postman, Insomnia, Bruno, OpenAPI or cURL into the workspace, and run_collection runs it with its tests (including pm.* scripts).
Full setup: Claude Code guide · MCP server overview
Updated 2026-11-19 · This page as Markdown · llms.txt