Building tools
Testing tools in chat
Every project ships with a built-in chat that uses the same MCP tool-calling loop as Claude Desktop and Cursor.
Why test in chat
Calling a REST endpoint directly tells you the API works. Testing in chat tells you whether the model can reliably figure out when and how to call it — which is the actual user experience.
What to look for
- Does the model pick the right tool when you ask in plain language?
- Does it fill in parameters correctly, or does it hallucinate values?
- Does it understand the response well enough to answer follow-up questions?
Iterating on descriptions
Tool descriptions are the single biggest lever for reliability. If the model misroutes, edit the description and re-test — no redeploy needed.