Coveo Hosted MCP Server

Coveo's Hosted MCP Server puts an enterprise search index behind the Model Context Protocol, so an LLM app can ground its answers in a company's own content without anyone building a custom integration first. It sits between the MCP client and Coveo's APIs, routing each tool call to the right endpoint and returning results the client can reason over.

The point of interest is that it reuses Coveo's existing relevance and permission model rather than re-implementing retrieval. Content that is already indexed and already permission-trimmed stays that way when an agent asks for it, which is the part most homegrown RAG pipelines get wrong.

The four tools.

• Search runs a full-text query through the Coveo Search API and returns ranked items — the general-purpose entry point for a research assistant that needs to explore and filter.
• Fetch retrieves one known item in full by its unique identifier, for cases like pulling up a specific policy document or technical manual on request.
• Passage Retrieval returns the most relevant sections of text rather than whole documents, which keeps context windows small and is the right tool when an agent needs supporting evidence from long contracts or manuals.
• Answer calls Coveo's Relevance Generative Answering to synthesise a natural-language response grounded in indexed content, suited to support deflection and knowledge-base bots.

The tools compose, so a support agent might answer directly while a legal assistant retrieves passages and reasons over them itself.

Access. It is generally available through ChatGPT Enterprise's Apps & Connectors directory and works with other major LLM apps and agent frameworks including Anthropic's Claude and AWS Bedrock. Authentication uses OAuth for secured content or anonymous API keys for public content. Usage falls under an existing Coveo AI-Relevance Platform subscription, with queries counting against consumption-based licensing rather than being billed separately.

Known limitations. Coveo documents three worth planning around: tool selection depends heavily on how clearly you write the tool descriptions in your configuration, so expect to iterate; there is no direct debugging capability yet; and facet-based filtering is not yet available.

A separate open-source reference implementation is published at coveo-labs/coveo-mcp-server for teams that would rather run the server themselves; it supports streamable HTTP, legacy SSE and stdio transports.