TL;DR
Mitzu's MCP server gives any MCP client, such as Claude, ChatGPT, Cursor or GitHub Copilot, one tool: Mitzu's analytics agent, which answers on your governed metrics. Unlike a raw warehouse MCP server, the assistant does not write SQL. The agent builds an analysis specification and Mitzu's deterministic engine generates the SQL, so the same question gets the same answer. Each person sees only what their Mitzu role allows, nobody needs a warehouse account, and the data stays in Snowflake, BigQuery, Databricks or your other warehouse.
Mitzu's MCP server is live. Claude, ChatGPT, Cursor, GitHub Copilot and any other assistant that supports the Model Context Protocol can now answer analytics questions from the metrics your data team defined in Mitzu. Revenue and orders, funnels, retention, segmentations and journeys: the same analyses you would run in the app, answered by Mitzu's analytics agent from inside the assistant people already use.
This post covers what the MCP server does, how it differs from connecting an assistant straight to the warehouse, and who it is for.
1. What Mitzu's MCP server does
Most MCP servers for data give an assistant a toolbox: list tables, describe a table, run a query. The assistant has to discover, select and chain those tools to reach an answer. Mitzu's MCP server gives the assistant one tool: Mitzu's analytics agent.
When a question arrives, the assistant passes it to the analytics agent, which runs the analysis and returns an answer built on your governed metrics, with the SQL behind it. No tool orchestration on the assistant's side and no skill chains to write.
The agent handles the same range of questions as in the app: a quick number or chart, a before-and-after comparison around a release or campaign, or a broad question such as why retention fell, which it answers by testing several angles and ranking what contributed. It uses the event, fact and dimension tables your data team exposes in Mitzu, so product events, orders and revenue can appear in one answer.
2. How it works
The trust differentiator in Mitzu is the deterministic query engine. The analytics agent does not write SQL. It assembles an analysis specification (funnel steps and conversion window, retention cohort, metric and breakdown) and the engine turns it into SQL using analytics methodology developed over years. Same specification, same SQL, same answer, every time.
The same mechanism runs behind the MCP server. When Claude or ChatGPT forwards a question, Mitzu's analytics agent picks it up, builds the analysis on your governed metrics, and returns the result. The calling assistant does not need to understand funnels, cohort bucketing or conversion windows; Mitzu's engine handles them.
A funnel with a correct conversion window, a retention chart with proper cohort bucketing, a segmentation that does not double-count users: these are not things the assistant has to get right. And because the definitions live in Mitzu, they are not tied to one AI vendor. Switch from Claude to ChatGPT and revenue is still revenue.
3. A governed analytics MCP server vs raw warehouse MCP servers
Warehouse vendors and the open-source community now offer MCP servers that connect an assistant straight to the warehouse. A Snowflake MCP server, for example, lets Claude or Cursor explore schemas and run SQL, and some route questions through the warehouse's own text-to-SQL service. Either way, a language model writes the SQL.
| Raw warehouse MCP server | Mitzu's analytics MCP server | |
|---|---|---|
| What the assistant gets | Tools to list tables, read schemas and run SQL | One tool: Mitzu's analytics agent |
| Who writes the SQL | A language model, for every question | Mitzu's deterministic engine, from the agent's analysis specification |
| Metric definitions | Inferred from table and column names, or a semantic model you maintain for it | The governed metrics your data team defined in Mitzu |
| Same question twice | Can produce different SQL and different numbers | Same specification, same SQL, same answer |
| Funnels and retention | Written by the model each time | Methodology built into the engine |
| Permissions | Whatever the warehouse role behind the connection allows | Each person's Mitzu role; nobody needs a warehouse account |
| Good for | Engineers exploring tables and drafting SQL | Anyone asking business questions on agreed numbers |
Connected straight to the warehouse, an assistant writes its own SQL for every question, so two people can get two different revenue numbers. Through Mitzu, every answer uses the definitions and permissions your data team set. Snowflake Cortex vs Mitzu goes deeper on the difference between warehouse text-to-SQL and a deterministic engine.
The two are not exclusive. If you are looking at a Snowflake MCP server so people can ask Claude about the business, decide whether you want the assistant writing SQL against Snowflake or answering from definitions your data team already approved. Mitzu runs on Snowflake and the other major warehouses, so engineers can keep a warehouse MCP server for exploration while everyone else gets governed answers through Mitzu's.
4. Permissions and data
- Permissions follow the person. Each person sees only the data their role in Mitzu allows, the same as in the app. Nobody needs a warehouse account.
- The warehouse stays private. The assistant receives the answer Mitzu returns, not your warehouse. You control whether query results go to the model, and your data is never used to train AI models.
- Same answer on every surface. The number in Claude matches the number in Mitzu's dashboards and in Slack, because every surface uses the definitions your data team approved.
5. Conversations and insights carry over to Mitzu
Each exchange through the MCP server is also a Mitzu conversation, with a link your team can open, continue or hand off in the app. It does not disappear when the chat session ends.
The insights the agent built along the way, such as funnels, segmentations and journeys, are saved in Mitzu. Open them in the app, add them to a dashboard, or return to them later.
6. Who it's for
Because anyone can ask from the assistant they already use, the MCP server reaches people across the company, not only those who open the analytics app every day.
- Customer success: checking usage on at-risk accounts while preparing a call, without switching tabs.
- Marketing: pulling campaign metrics into a planning doc, from the tool where the doc is being written.
- Product managers: getting post-release usage numbers for a deck, directly from chat.
- Engineers: prioritizing bug fixes with product data, from the IDE where the code lives.
- Data teams: defining metrics once and letting people ask from Claude, ChatGPT or Cursor, while the definitions, permissions and results stay in Mitzu.
This is the same idea as Mitzu's Slack app, where people mention @mitzu in a channel and the agent answers in the thread. The MCP server extends it to any assistant that speaks the Model Context Protocol.
7. Compatible assistants
Mitzu's MCP server works with Claude, ChatGPT, Cursor and GitHub Copilot, and any other assistant that supports MCP, including agents your team builds. Setup instructions are in the Mitzu MCP documentation.
8. Warehouse readiness: the prerequisite
The MCP server runs on the same foundation as the rest of Mitzu: event and business data in a cloud data warehouse such as Snowflake, BigQuery, Databricks, Redshift or ClickHouse (see all supported warehouses), with metrics defined in Mitzu. Queries run in your warehouse and nothing is copied out. If your event data still sits in a third-party vendor silo without warehouse export, Mitzu is not the right fit yet.
Frequently asked questions
What is Mitzu's MCP server?
MCP, the Model Context Protocol, lets AI assistants call external tools. Mitzu's MCP server gives an assistant one tool: Mitzu's analytics agent. The assistant asks it a question and gets back an answer built on your governed metrics, with the SQL behind it.
How is an analytics MCP server different from a Snowflake MCP server?
A warehouse MCP server, such as one for Snowflake, gives the assistant access to tables and lets a language model write SQL against them. Mitzu's analytics MCP server gives the assistant an analytics agent that answers from the metric definitions and permissions your data team set, with SQL generated by a deterministic engine. The same question gets the same answer, in the assistant and in Mitzu.
Does the calling assistant (Claude, ChatGPT, Cursor) write the SQL?
No. Neither the calling assistant nor Mitzu's analytics agent writes SQL. The agent assembles an analysis specification, such as funnel steps, retention parameters or a breakdown, and Mitzu's deterministic query engine turns it into SQL. The same specification always produces the same SQL.
The calling assistant only asks the question.
Do Mitzu permissions apply in the assistant?
Yes. Each person sees only the data their role allows, the same as in Mitzu. Nobody needs a warehouse account.
Which assistants are compatible?
Claude, ChatGPT, Cursor and GitHub Copilot, and any other assistant that supports MCP. See the Mitzu MCP documentation for setup instructions per assistant.
Does this require a specific Mitzu plan?
See Mitzu pricing for which plans include the MCP server and the analytics agent.
Get started
Connect your assistant to Mitzu's MCP server, point it at your workspace, and start asking questions. Full instructions are in the Mitzu MCP documentation, and the MCP and Slack page has an overview.
Related reading
- ChatGPT vs an AI analytics agent: CSVs, warehouse connections and MCP
- What is agentic analytics?
- Why analytics agents need a semantic layer
- How AI agents query your data warehouse
- Planning mode: review the agent's plan before it runs
