Model Context Protocol

Model Context Protocol (MCP) is an open protocol for connecting LLM applications to external data and tools. The 2026-07-28 specification ↗ distinguishes the host application, its clients and capability-providing servers.^[1] MCP is not itself a model, retrieval-quality measure or decision policy.

An AI Agent can connect through MCP; RAG can use retrieved material as context. MCP support alone establishes neither accurate retrieval, unrestricted access nor blanket permission to execute actions.

Resources supply readable context, prompts supply structured messages or workflows, and tools expose callable functions.^[1] In a fictional wiki integration, article content, a comparison template and a search function serve different roles. Knowing an input schema does not establish that the requested target or authority is appropriate.

Servers need not implement every feature. Separate discovery, proposed invocation, actual execution and result verification to distinguish connectivity failures from decision errors.

The release announcement ↗, checked on 20 October 2026, describes a stateless core, per-request version and capability information, and an optional client call to server/discover. This revision replaces reliance on the earlier initialize/initialized exchange and protocol session identifier.^[2] Older examples should not be applied indiscriminately.

Multi Round-Trip Requests (MRTR) return requests for additional input and let the client retry the original call with answers. Check support for extensions such as Tasks separately.^[2] Protocol statelessness does not eliminate persistent data, long-running work or side effects. Record the supported protocol and SDK versions together.

This is an ORBiS Tree design example, not a description of an available MCP server. For a task that compares two official sources and saves a draft, retain source URLs, observation times and document identifiers; the save result should identify the draft and version. If the response is lost, provide a way to inspect saved state before blindly repeating the write.

Successful search with read access does not establish publishing authority. A retrieved instruction to publish is not a new user request. Verify the requested target, a single saved draft and retained evidence before counting completion. See AI Agent Oversight and Prompt Injection.

MCP Authorization ↗ covers HTTP authorization and validation that access tokens were issued for the intended server.^[3] The security guidance ↗ discusses risks of passing tokens to other APIs without appropriate validation.^[4] Copying credentials into model context is not a substitute.

Valid authentication concerns identity and permitted access. Whether a tool-result instruction is authoritative, and what change the user currently requested, are separate questions. Consult the linked articles for execution boundaries and outcome checks.

A shared protocol does not make server capabilities, supported versions or permissions identical. Schema-valid calls can still target the wrong resource or return inaccurate material, requiring semantic outcome checks. ORBiS Tree proposes testing unsupported versions, invalid arguments, expired authority, additional-input requests, cancellation, lost responses and external instructions alongside ordinary calls. Record connection success, task completion, incorrect writes and retry costs separately. No one measure represents the whole integration.

Agentic Coding concerns repository changes, Computer Use concerns interfaces, and AI Agent concerns action selection. MCP is one connectivity mechanism. Compare it with available APIs according to required capabilities and verifiable outcomes.