MCP Engineering

Design a capability surface that a host/model can discover and use correctly, and a server can enforce under retries, races and reconnects. Ordinary API correctness still applies; MCP adds negotiated host capabilities and model interpretation.

Workflow

1. Identify the real caller/host, supported protocol/SDK versions, transport, negotiated capabilities and authority boundary. Test support rather than assuming sampling/resources/prompts are exposed uniformly.
2. Classify capabilities by use: tools for model-selected operations (including useful reads); resources for addressable context where hosts support discovery/attachment; prompts for reusable workflows; host sampling only when negotiated and suitable.
3. Write schemas/descriptions and bounded outputs around user intent. Name required permissions, side effects, error/recovery classes, pagination and compatibility.
4. State durable idempotency/concurrency/atomicity guarantees for mutations and observable external effects. Trace logical operation separately from attempts.
5. Implement and verify protocol/contract behavior deterministically; replay captured calls with explicit volatile-field binding/comparison. Frozen calls do not evaluate how a model interprets descriptions.
6. Use real model/host tasks where tool selection or interpretation is uncertain. Record task outcomes, model/host configuration, failures and denominator; no universal pass quota.
7. Review evidence, supported limitations and recovery. Clean results are allowed after a scoped sweep; do not require critic disagreement.

An external inference SDK in a server is a product/consent/data/credential choice, not automatically a blocker. Distinguish server-owned inference from borrowing host inference; disclose the trust/cost boundary and evaluate negotiated alternatives.

Optional references

Primitive selection, tool contracts, errors, idempotency, bounded output, trust, transport, resources/prompts/sampling, composition, versioning, telemetry, tests, smells.