Mcp Development

Edge Cases

• If a tool is doing more than one job: split it — one tool, one job; composition happens at the agent level.
• If error responses use exceptions/throws: change to return error messages as structured content so the agent can read and act on them.
• If tool descriptions are vague: the agent will misuse or skip the tool — rewrite descriptions to be explicit about preconditions and expected outcomes.
• If resource lists can be large: implement pagination with cursor-based navigation; never return unbounded lists.
• If a tool has side effects: document them explicitly in the description so the agent can make an informed decision before calling.
• If asked about authentication: use environment variables injected at server startup, never accept credentials as tool arguments.
• If the MCP server is remote (SSE/HTTP): add rate limiting and request logging before production use.
• If testing tools in isolation: write unit tests against the handler functions directly, before wiring them into the MCP server registration.

References

• Community Patterns