hi-donwi MCP Builder
Overview
Expose tools to agents through narrow, well-described, permission-aware MCP interfaces.
When to use
- Building an MCP server for a service, database, internal system, or workflow.
- Reviewing MCP tool schemas, auth, permissions, or safety.
- Turning an external API into agent-usable tools.
Process
- Define use cases. Start from agent tasks, not raw API endpoints.
- Design small tools. Each tool should have a clear action, typed inputs, predictable output, and explicit side effects.
- Model auth and permissions. Least privilege, scoped credentials, no secret echoing, and clear approval boundaries for mutating actions.
- Validate schemas. Use structured inputs/outputs, enum constraints, limits, and helpful error messages.
- Handle failures. Timeouts, retries, rate limits, idempotency, partial failure, and user-safe recovery paths.
- Evaluate with agent tasks. Test representative prompts and verify the agent can choose the right tool without guessing.
- Document setup. Transport, environment variables, credentials, local dev, and verification commands.
Red flags
- One giant
run_any_queryorcall_any_endpointtool exposed by default. - Tools with vague names or unbounded string input.
- Mutating actions without preview/confirmation or idempotency.
- Returning raw huge API payloads instead of agent-usable summaries.
- Secrets in logs, prompts, tool descriptions, or error messages.
Verification
- Tool schemas are typed and constrained.
- Auth, permissions, and side effects are explicit.
- Representative agent prompts succeed against local or test MCP server.