Model Context Protocol
Definition
Model Context Protocol (MCP) is an open protocol that standardises how an LLM-based agent connects to external systems — data sources, tools, and services — so that a capability is exposed once and consumed by any compliant client, rather than reimplemented per agent.
In a search context it matters because it fixes the boundary between the agent and the engine: the agent does not hold engine credentials or construct privileged calls directly, it asks an MCP server which exposes a defined surface.
Why It Shows Up in Search Systems
Search stacks are an awkward fit for direct LLM access. The engine speaks a query DSL, holds production data, and is operationally sensitive — three reasons not to hand a model a raw connection. An MCP server in front of it gives:
- A stable surface. Tools are declared with typed inputs, so the agent’s access is enumerable rather than open-ended.
- A security boundary. The OpenSearch Relevance Agent routes every agent call through the OpenSearch MCP server, described there as a secure translator between the AI and the search engine.
- Reuse across clients. The same server serves whatever agent or IDE speaks the protocol.
In This Vault
- OpenSearch Relevance Agent — all three of its agents communicate with the engine exclusively over the OpenSearch MCP server (
opensearch-mcp-server-py); multi-platform data source connectivity over MCP is on its roadmap. - Elasticsearch Relevance Studio — also exposes its capabilities over MCP.
Related Concepts
- Agentic Search — the pattern MCP is plumbing for
- Agentic Memory — the other half of what an agent needs beyond tool access
- Search Observability — what an enumerable tool surface makes measurable
Related Tools
Articles
- Introducing OpenSearch Relevance Agent — MCP as the mandatory abstraction between agents and engine