MCP Server Development

Quick start

• Treat an MCP server as an AI-user interface (outcomes), not an endpoint wrapper.
• Keep the server small: one bounded context, ~5–15 tools max.
• Design tool contracts first: names, descriptions, flat input schemas, concise outputs.
• Implement with an SDK (Python or TypeScript), validate inputs, and return structured errors.
• Test for correctness, safety, and token efficiency.

Workflow

1. Define the server’s bounded context and user outcomes (see references/quick-reference.md).
2. Draft tool contracts (names, descriptions, schemas, outputs) (see references/tool-design.md).
3. Choose an implementation library and transport (see references/python-libraries.md or references/typescript-libraries.md).
4. Implement validation, errors, and pagination conventions (see references/protocol-and-schemas.md).
5. Add security guardrails (see references/security.md) and observability (see references/observability.md).
6. Add tests (see references/testing.md) and set a versioning policy (see references/versioning.md).