A hosted MCP server for your database and spreadsheets
Connect a data source, publish endpoints, and every one of them becomes a tool your AI assistant can call — on a hosted MCP server with OAuth 2.1, at a URL that belongs to your workspace. There is nothing to deploy and nothing to maintain.
What an MCP server actually is
The Model Context Protocol is an open standard for connecting AI assistants to things outside their own context window — tools they can call, and data they can read or change. An MCP server is simply a service that advertises a set of those tools. When you connect an assistant to one, it can see what is available, what arguments each tool takes, and what it gets back.
Without something like MCP, giving an assistant access to your data means pasting it into the conversation. That falls apart quickly: exports go stale the moment they are made, large datasets do not fit, and nothing can be written back. With an MCP server the assistant queries live data on demand and only pulls the rows it actually needs.
The catch is normally that somebody has to build and host the server, define every tool schema by hand, and then keep those definitions in step with the underlying data as it changes. That is the part Dataloom removes.
Why a hosted MCP server beats running your own
Writing an MCP server is not hard. Operating one, securing it, and keeping its tool definitions honest as your schema changes is the part that costs you.
Tools stay in sync automatically
Publish an endpoint and it appears as a tool roughly a minute later. Unpublish it and it disappears. Connected assistants pick up the change without anyone reconnecting or editing a config file.
OAuth 2.1, not pasted keys
Authorisation goes through a proper OAuth flow. You sign in and approve a workspace; the assistant gets a scoped token. No long-lived secret to copy between machines or leak into a transcript.
Typed schemas from your real data
Each tool advertises the actual columns your endpoint returns, so the assistant filters and sorts on fields that genuinely exist rather than plausible-sounding guesses.
Per-assistant revocation
Every connected assistant can be revoked on its own from your workspace. Removing access from one does not disturb the others or force a rotation.
How to set it up
- 1
Publish at least one endpoint
Connect a data source, build an endpoint and publish it. Give each endpoint a clear description — the assistant reads it to decide when the tool is relevant, so a good description makes a noticeable difference to how well it behaves.
- 2
Enable MCP for the workspace
Open MCP in the Dataloom sidebar and switch it on, then copy the server URL shown there. Each workspace has its own, so you can connect several workspaces to the same assistant at once. Only workspace managers can enable or manage MCP.
- 3
Add the connector to your assistant
Paste the URL into your assistant as a custom connector. You will be sent to Dataloom to sign in and approve access, and from then on the connection is remembered.
The full setup guide covers claude.ai, Claude Desktop, Claude Code, Cursor, VS Code and ChatGPT in detail.
// mcp-config.json
{
"mcpServers": {
"dataloom": {
"url": "https://mcp.dataloom.cloud/w/your-workspace",
"transport": "http"
}
}
}What the assistant can and cannot do
Handing an AI assistant a database connection is a bad idea, and Dataloom does not do that. The assistant never sees your credentials or connection string. It sees a list of endpoints you chose to publish, and nothing else about the source behind them.
Every call still travels through the same gateway your other API traffic uses, so your auth rules, rate limits and per-endpoint configuration all apply unchanged. Requests appear in your logs and analytics like any other request. If you published read-only endpoints, reads are all that is possible — writes are only available where you deliberately published a write endpoint and the underlying credentials allow it.
- Only published endpoints are visible as tools
- No credentials or connection strings are ever exposed
- OAuth 2.1 authorisation instead of shared API keys
- Your existing rate limits and auth rules still apply
- Every call is logged like any other API request
- Any connected assistant can be revoked on its own
Serve any of these to your AI agent
The MCP server is the same whatever sits behind it. Connect one of seven sources and publish endpoints from it.
MCP questions, answered
What is an MCP server?
MCP (Model Context Protocol) is an open standard for giving AI assistants access to external tools and data. An MCP server exposes a set of tools an assistant can call. Dataloom runs one for your workspace and fills it with the endpoints you publish.
Is the MCP server really included on the free plan?
Yes. The hosted MCP server is included on every plan, including Free. You do not need a paid plan or a credit card to connect an AI assistant to your data.
Which AI assistants does it work with?
Any MCP-compatible client. That includes claude.ai custom connectors, Claude Desktop, Claude Code, Cursor, VS Code and ChatGPT.
Do I have to paste an API key into my AI assistant?
No. Connections use OAuth 2.1 — you sign in to Dataloom and approve the workspace, and the assistant receives a scoped token. There is no long-lived key to copy around or accidentally leak into a chat log.
Can an AI assistant see data I have not published?
No. The only things exposed as tools are the endpoints you have explicitly published. Anything you have not published is not reachable, and every call still runs against the credentials and permissions of the underlying data source.
How do I disconnect an assistant?
Each connected assistant can be revoked individually from your Dataloom workspace, without affecting the others or requiring you to rotate anything.