What each platform's MCP server exposes to agents

An MCP server is now close to a requirement for any product that wants to give context to AI agents. That makes the presence of one a weak signal. The stronger signal is what an agent can ask for and what it gets back. This memo lists what each vendor documents on its public pages.

Euno says it surfaces metadata-powered context to AI agents through MCP and describes four layers in that context: semantic search and lineage insights, active metadata tags, usage insights and personas covering role, task and access scope. Its pricing page lists Context API calls of up to 500K or 1M by tier. Euno also says it computes its context graph ahead of time rather than gathering context through MCP at query time.

Source: Euno: Context for AI Agents

Source: Euno pricing

Source: Euno blog: AlphaSense x Euno at dbt Summit

Atlan serves context through SQL, APIs, an SDK and the Atlan MCP server, and lists evals and traces in the same step. Its homepage quotes Workday on its use of the MCP server.

Source: Atlan homepage

DataHub lists MCP alongside an Agent Context Kit, Skills, APIs and SDKs, and names Claude, Cursor, Snowflake Cortex, Databricks Genie, CrewAI, LangChain and Agent Development Kit as agents that can use the context. Of the platforms we read, it documents the widest set of agent interfaces.

Source: DataHub Cloud context platform

Collibra's MCP Server ships 25+ governance tools out of the box, including asset discovery, lineage exploration and classification management. Collibra describes it as token-efficient and deployable anywhere.

Source: Collibra MCP Server

Alation's Ontologies can be published as an MCP server. Ontologies was in early access as of Alation's 17 September 2026 release.

Source: Alation AIOS: Ontologies

Source: Alation press release: AIOS expansion with six products

Select Star lists an MCP Server for Data, described as a single API, alongside Ask AI.

Source: Select Star homepage

AtScale documents its MCP server in the most detail of any vendor we read. It names the tools: list_models, explore_columns, focus_columns, get_sml_skills, run_query and get_outbound_queries. It names the clients: Claude, ChatGPT, Microsoft Copilot, Grok, Codex, LangGraph and Google ADK. Each agent gets its own OAuth identity, with row- and column-level security and a log of every call.

Source: AtScale: semantic context for AI via MCP

Cube includes an MCP server for AI agents in its free plan and describes governed data reaching agents through MCP and APIs; its Analytics Chat API is described as agent-to-agent capable via MCP.

Source: Cube homepage

Source: Cube pricing

Honeydew supports an AI analyst use case, but we did not confirm an MCP server on the pages we read.

Source: Honeydew pricing

Modus says agents such as Claude and Cursor call it over MCP or API and get back permission-checked context. Stardog's MCP server appears only in a secondary source in our reads, and Jedify's release does not mention MCP.

Source: Modus homepage

Source: GlobeNewswire: Jedify raises $24 million Series A

Read side by side, the documentation falls into three patterns. Governance-first servers, such as Collibra's, expose tools for discovering assets and checking classifications, so an agent can ask whether something is allowed before it acts. Metric-first servers, such as AtScale's and Cube's, expose certified measures and a query tool, so an agent can ask for revenue by name and get the governed number. Context-first servers, such as Euno's, Atlan's and DataHub's, return a slice of a wider graph: definitions, lineage, trust labels and usage around the asset the agent is about to use.

None of these patterns is wrong. They answer different failure modes, and the documentation alone will not tell you which one your agents hit most often. Your benchmark questions will. If your agents compute metrics inconsistently, a metric-first server addresses that directly. If they pick untrusted or sensitive assets, governance-first or context-first servers do more. Many teams will end up connecting an agent to two servers, which makes the overlap between them worth checking.

Ask four questions of every server. Which tools does it expose, by name? What does a response contain for one of your tables: description only, or lineage, certification and sensitivity too? How is the context scoped to the person the agent acts for? How long after an upstream change does the response change?

Then connect the agent you will actually use and record the raw responses for ten real questions. Those records will tell you more than any feature table, including ours. Our agent access criterion gives the highest scores to servers that are generally available, documented in detail and reachable from several agent clients.