Kaivo MCP Server
The Kaivo MCP server lets AI tools query your data warehouse.
Written By Lauri Raivio
Last updated About 3 hours ago
The Kaivo MCP server lets AI tools query your data warehouse. You connect a tool once, then ask questions about your business in plain language and get answers from your own data.
MCP stands for Model Context Protocol. It's an open standard for connecting AI assistants to outside systems, supported by Claude, ChatGPT and a growing number of other tools. Rather than each AI vendor building its own Kaivo integration, we run one server that any of them can connect to. In practice you paste an address into your AI tool, sign in to Kaivo, and choose which workspace it may read.
What you can do with it
Ask the questions you'd otherwise put to an analyst or work out in a BI tool:
"How did sales develop over the last six months?"
"Which customers haven't ordered since spring?"
"Compare campaign spend to revenue by channel last quarter."
"Where does the customer email address live in our warehouse?"
You don't need to know your schema or write SQL. The AI explores your warehouse, finds the tables that hold the answer, reads their column descriptions, and writes the query itself.
It can also save a result back into your warehouse as a new table, so a BI tool or a later question can pick up where it left off. Those saved tables go into a dedicated dataset called kaivo_ai_outputs, which is the only place anything can be written.
Which AI tools can connect
Claude works fully, on claude.ai, Claude Desktop and Claude Code. See Connecting Claude to Kaivo.
ChatGPT works on Business, Enterprise and Edu plans, where an administrator enables Developer Mode. On Plus and Pro, ChatGPT can only use the search and fetch tools, which is enough to find tables and read their structure but not to run queries. Connectors aren't available on the free plan.
Microsoft Copilot can connect through a Power Platform custom connector. This is more involved than the others, so talk to us if you need it.
Any other tool that supports remote MCP servers over HTTP should work. If you're trying one that doesn't, let us know.
The tools
An MCP server offers the AI a set of tools it can call. Ours has eight. Most AI tools show you this list and let you turn individual tools on or off.
Seven of the eight only read. The one that writes can only write to kaivo_ai_outputs, and it's labelled so AI tools can ask first.
Each tool is labelled so your AI tool knows which ones are safe to run without checking with you first. In Claude, for example, read-only tools run as part of answering your question, while saving a table prompts you to confirm.
Limits
Every query is checked before it runs. If it would break one of these, it's refused and the AI is told why, so it can try a narrower question instead.
1 GB scanned per query. Ask for a date range or a summary rather than scanning years of history.
1000 rows returned per query. If more rows matched, the AI is told the result was cut short so it can tell you.
120 requests a minute per workspace.
Reads only your workspace. Nothing can reach another workspace or another customer's data.
Internal tables are hidden. The staging tables our pipelines use to build your warehouse aren't visible.
Existing data can't be changed or deleted. The only writes go to
kaivo_ai_outputs.
What the AI can see and where your data goes
The AI sees exactly what you see. When you connect, you pick one Kaivo workspace, and the AI can read the datasets in it: the same ones you can already open in Kaivo. It can't reach a workspace you don't have access to, and the internal tables our pipelines use stay hidden. Because access follows your sign-in, even a session that has been fed misleading instructions can only reach what you can reach.
The AI queries your warehouse directly. Kaivo already keeps your data in the warehouse we build and maintain for you. Connecting an AI tool doesn't add a second copy of it: nothing is exported, indexed, or moved anywhere else to make this work. Queries run in the EU against your live data, and there are no keys for you to look after.
What comes back is the result of your question. The AI writes a query, we run it, and the results go back to the AI, which means they reach your AI vendor's servers rather than your own computer. If you ask for a list of customers, that list is what gets sent. Each query returns at most 1000 rows, so it is never a full copy of a table.
You choose the AI tool, so your own agreement with that vendor is the one that applies. Kaivo never brings in an AI of its own. Whatever your agreement with them says about training and retention covers the results the AI receives.
Every question leaves an audit trail: who asked, what was queried, and how much data it scanned.
Who can connect, and how to take access away
Anyone in your workspace can connect an AI tool to it. Each person signs in with their own Kaivo account and chooses their own workspace, even when an administrator added the connector for the whole organisation. Your name is what appears against your connection.
The Members & access page in Kaivo lists everything currently connected, who connected it, and when. You can disconnect any of them from there, and access ends within a few minutes. Disconnecting one doesn't affect the others.
Removing a connector inside the AI tool itself only stops that tool from using it. It doesn't withdraw access to your data. To do that, disconnect it in Kaivo.
Getting started
We have separate guides to getting started depending on your AI tool:
If you don’t see the tool you are using listed above, don’t hesitate to contact us for more help. Kaivo can work with any Any MCP client that can reach a remote server over HTTP and sign in with OAuth, so your tool is more than likely already supported.
Questions about any of this? Email support@kaivo.io.