AI Summary: MCP is an open protocol for connecting LLM applications with external context and tools through JSON-RPC 2.0. A website is not MCP-ready merely because it has HTML or an API: it needs an intentional server boundary, typed capabilities, consent, authorization, privacy controls, and safe handling for every tool operation.
What MCP adds
The Model Context Protocol specification connects LLM hosts to external data and tools through JSON-RPC 2.0. It separates hosts, clients, and servers and negotiates capabilities during initialization. [1]
Servers expose resources, prompts, or tools: resources provide context, prompts provide workflows, and tools request operations. Keep these contracts distinct rather than hiding read and mutate behavior behind one endpoint.
Design the boundary first
Inventory what an assistant may read, suggest, or mutate. Give resources stable identifiers and access rules; give tools narrow schemas, output and error shapes, and idempotency expectations. Require authorization at the server boundary and confirmation for destructive actions.
The MCP specification says users must understand and consent to data access and operations. It also treats tools as arbitrary code-execution paths that need caution and explicit consent. [1] Do not expose private customer data because a client can guess a URI. Do not rely on a tool description as a security permission. Enforce identity, scope, audit logging, rate limits, and revocation in the server.
| Capability | Readiness evidence |
|---|---|
| Resource | Stable URI, owner, access rule, and representative response |
| Prompt | Versioned template with inputs and intended user action |
| Tool | Narrow schema, authorization check, error contract, and audit trail |
| Transport | JSON-RPC handshake, capability negotiation, timeout, and retry policy |
| Consent | Human-readable approval before sensitive access or mutation |
Verification checklist
Connect a test host to a non-production server and capture initialization capabilities. Test allowed and unauthorized resources, malformed input, duplicate mutation, and revoked credentials. Verify audit logs record actor, tool, target, result, and failure without secrets.
MCP readiness is a controlled integration surface, not an HTML badge. Publish only capabilities you can authorize, observe, and safely withdraw.