FDE Times: Build a fake MCP server that injects timeouts, 429s and schema changes to test agents offline

Demo agents look fine until a customer CRM rate-limits, hangs for 30 seconds, or renames a field. FDE Times’ 4 November guide builds a fake MCP server (Python + FastMCP) that exposes the same tool names as production but fails on a fixed FAULTS script—so tests need no network and stay repeatable.

Under the MCP spec, tool failures travel two channels: JSON-RPC protocol errors (e.g. unknown tool, -32602) and execution errors with isError: true. The minimum fault set covers a 429 with retry-after, a spend-cap 429 without retry-after, a hung tool cancelled by timeout, a misspelled tool name, and schema drift (e.g. order_id → orderId, number → string) that returns success on neither error channel.

The arithmetic trap: Claude SDKs already retry transient errors twice by default; if your agent loop also retries three times, one 429 can become 12 requests on an overloaded API. Put retry logic in one layer, wait on retry-after when present, and stop instead of looping forever on a spend-cap style failure.

In week one without staging access, match the customer’s rate-limit docs and build the suite in parallel; each real incident should add another FAULTS line. On a CV, cite the number of scenarios and bugs caught before production—not just “built agents with MCP.”