Skip to content

MCP Server Overview

Mention exposes a Model Context Protocol (MCP) server so external agents — Claude Desktop, IDE-integrated assistants, custom internal tools — can use your organization's curated knowledge as a backend. An MCP client connects to the server with an Agent API key and can then list Audiences, read Articles and SOPs, and ask grounded questions exactly as a logged-in member could, without having to log in.

This page covers what the MCP server is and how to connect to it. For the tool surface itself, see MCP Tools Reference. For the credentials, see Agent API Keys.

What MCP is

The Model Context Protocol is an open protocol for connecting AI agents to external tools and data sources. An MCP server exposes a set of tools the agent can call and resources it can read; an MCP client (the agent) connects to the server, discovers the tools, and uses them as part of its reasoning.

Mention's MCP server is one such server. It lets agents bring your curated Glossary and Playbook content into their context — useful when you want an agent to answer questions about your team's terminology and processes the same way Mention does internally.

Why connect via MCP

The MCP server is the right surface when you want:

  • An external agent to answer questions grounded in your organization's source material without re-implementing retrieval, Article generation, or SOP generation.
  • A consistent, curated knowledge layer across multiple agents — internal chatbots, IDE assistants, voice agents — without copy-pasting your content into prompts.
  • Access scoped to specific Audiences, with credentials you can rotate or revoke independently per consumer.

If you want to embed Mention content in your own UI, or build something more bespoke than what an MCP client gives you, the same data is reachable through the underlying API. The MCP server is the lowest-effort path for the common case.

Authentication

The MCP server authenticates with an Agent API key, passed as a Bearer token in the Authorization header. Each Agent is scoped to one or more Audiences — the server only exposes content from Audiences the Agent has been granted.

To get a key, an admin creates an Agent in your organization and grants it the Audiences the consumer should see. See Agent API Keys for the developer-side details on using the key, and Agents for the admin-side flow of creating and managing them.

How to connect

Most MCP clients (Claude Desktop, IDE extensions, etc.) take a server URL plus headers. Configure them with:

  • URL — https://mcp.mentionai.app/mcp
  • Header — Authorization: Bearer <agent-api-key>

A typical client config looks like:

"mention": {
  "url": "https://mcp.mentionai.app/mcp",
  "headers": {
    "Authorization": "Bearer ${MENTION_TOKEN}"
  }
}

Set MENTION_TOKEN as an environment variable in the shell that launches your MCP client (don't paste the raw key into shared config files). When an admin creates or rotates an Agent, Mention shows the token alongside a ready-to-paste version of this snippet.

Once connected, the client discovers the available tools automatically. There is no separate handshake or session step.

What the server exposes

The MCP server exposes six tools, in roughly the order an agent would use them:

  1. list_audiences — discover which Audiences the Agent has access to.
  2. read_glossary — read the Audience's Glossary, including concept IDs and prerequisites.
  3. read_article — read the full Article for a specific Concept.
  4. list_processes — list the Processes in an Audience's Playbook.
  5. read_sop — read the full SOP for a specific Process.
  6. ask — ask a synthesized cross-cutting question after reading the relevant material.

Each tool is described in MCP Tools Reference, including when to prefer one over another.

Generation latency

The first time an agent reads an Article that hasn't been generated yet, read_article triggers generation and polls until it's ready. Articles take up to roughly a minute; SOPs up to roughly five minutes. After the first read, content is cached and returns immediately on subsequent calls.

Plan around this in your client: a first-time read_article is occasionally slow; subsequent reads are not.