AI assistants
Connect Claude, ChatGPT or another assistant to a space so it can read your data, and build and publish on it.
An AI assistant you already use can connect to a Flitch space. Once connected it can list your datasets and dashboards, answer questions of your data where it sits, and, if you allow it, build a dashboard, report or app from a brief and publish it for a link. Flitch holds the connections and the schema, so the assistant never needs credentials for the systems behind your data.
Assistant connections are available on every plan. A build started from an assistant is paid for with the space's credits, like one started in Flitch.
Connect an assistant
Flitch runs a remote MCP server at:
https://app.flitch.io/api/mcpAny assistant that supports remote MCP servers with OAuth can connect to it. In Claude, add it under Settings → Connectors as a custom connector. In ChatGPT, add it as a connector from Settings → Connectors. Other assistants have an equivalent place to add an MCP server by URL.
Adding the server starts a sign-in with your Flitch account:
- Choose a space. A connection reaches one space. To use another, connect again and pick it.
- Approve what it may do. The consent screen lists each permission the assistant asked for. Allow or deny it as a whole.
- Return to the assistant. It is now connected, and stays connected until you disconnect it.
Anything the assistant reads from Flitch is shared with that assistant's provider under its own terms.
What an assistant can do
Each permission is granted at consent and enforced on every call.
| Permission | What it allows |
|---|---|
| Read your data | List datasets and their columns, run read-only SQL over up to four datasets at once, and see which sources are connected and when each last synced. |
| Read your dashboards, reports and apps | Find what exists in the space, what each one shows, which datasets it reads, and its published link. |
| Build and publish dashboards, reports and apps | Start a build from a brief and the datasets it should read, follow its progress, and publish the result. |
Reading
A query returns at most 200 rows, and a dataset over 50,000 rows is refused rather than read in part, so an assistant is steered to aggregate in SQL rather than add rows up itself. Column names are exact, so a well-behaved assistant describes a dataset before querying it.
Building
A build needs every dataset it should read named up front. With none named it does not start, and the assistant is handed the datasets to choose from. The build runs on Flitch and usually takes a few minutes; the assistant polls for progress and reports back. The response to a build request says what recent builds in the space have cost, so the assistant can check with you before starting one.
Publishing
Publishing returns a link on view.flitch.io and takes an access type: public for anyone with the link, login for any signed-in Flitch user, or team for members of the space. Publishing a dashboard that reads row-restricted data as public turns that restriction off for the published copy, so the assistant is told to confirm the access type with you first. Publishing again after an edit updates the live copy.
What it cannot do
A connection cannot change a dataset, a source or a setting, and it cannot write rows. Those stay in Flitch, or go through the API from your own code.
Manage connections
Open Settings → AI assistants.
- Your connections lists every assistant you have connected, in every space you belong to, with what each may do and when it was last used. Disconnect one and it loses access straight away.
- Other members' connections lists assistants other members have connected to the current space. An admin of the space can disconnect any of them.
- Allow AI assistants in this space is the switch for the whole space. Off, no assistant can reach it, including ones already connected; on again, they work as before.
API keys and assistants
An assistant connects with your sign-in and is open on every plan. API keys, for your own code, remain a Pro and Team capability. An assistant can also be given an API key as a bearer token instead of signing in, in which case the key's plan and scopes apply. See authentication.