Testing

Test MCP endpoints before production

A preflight checklist for generated MCP endpoints: initialize, tools/list, schema validation, auth failure, safe execution, and logging.

Jun 25, 202610 min readIntermediate

Steps

Initialize the endpoint

Start with the protocol handshake. If initialize fails, do not debug individual tools yet.

Example

curl -sS -X POST https://astrail.dev/api/mcp/SERVER_ID \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}'

Inspect the visible tools

tools/list should show only the tools your public policy allows. Check names, descriptions, input schemas, and auth annotations.

Example

curl -sS -X POST https://astrail.dev/api/mcp/SERVER_ID \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'

Send a known-bad argument

A production MCP endpoint should reject invalid input before making an upstream call. This catches loose JSON schema mapping and unclear required fields.

Example

curl -sS -X POST https://astrail.dev/api/mcp/SERVER_ID \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 3,
    "method": "tools/call",
    "params": { "name": "get_pet_by_id", "arguments": { "petId": "" } }
  }'

Run one read-only happy path

Pick a GET route with deterministic output. The first production smoke should prove end-to-end execution without modifying upstream state.

Review runtime evidence

Check trace id, upstream status, latency, execution mode, and redaction. The goal is not just a green response; it is a debuggable response.

Production checks

initialize returns server metadata and capabilities.

tools/list excludes private tools from public surfaces.

Invalid arguments fail locally with a clear validation error.

Auth-required tools fail closed when credentials are absent.

Runtime logs redact credentials and include enough context to debug.

FAQ

What is the minimum MCP smoke test?

Run initialize, tools/list, one invalid tools/call, one auth failure, and one read-only happy path.

Should smoke tests call destructive endpoints?

Only in a staging environment with test data. Production launch checks should prefer read-only paths unless you have explicit rollback behavior.