Airtable

A personal access token needs two settings, not one: the scopes, and then each base explicitly granted to it. A token with every scope and no base granted answers 200 with an empty list rather than an error, which is the failure that looks like an empty table. Worth testing the connection before building anything on top of it.

Is there an MCP for Airtable?

Yes. Airtable hosts its own MCP server at https://mcp.airtable.com/mcp, available on all plans, and it follows the signed-in user's base role: owners and creators can change schema and records, editors can change records only. Tulina reaches the same bases through its own MCP server, on the same link that carries the rest of your stack, so one conversation can read a table, check a lead against your CRM and post the result to Slack.

How do you connect Airtable to Claude and ChatGPT?

Through Tulina's own MCP server, in three steps:

1. At airtable.com/create/tokens, create a personal access token with the data.records, data.recordComments and schema.bases scopes, then add each base it should see under Access.
2. Paste the token into Tulina's Airtable connector, for yourself or for the organisation. It is stored in the vault and never shown to the model.
3. Add https://mcp.tulina.ai/mcp to Claude or ChatGPT and sign in with your Tulina account. The MCP page has the same link.

Several tokens can be connected as named accounts, for example one per client, and a request names the one it is for.

Tulina vs Airtable's official MCP server

Both reach the same bases. They differ in how you sign in and in what each one exposes.

If Airtable is the only system the model needs, or you want it to build interfaces and automations, Airtable's own server is the better choice. Tulina earns its place when a table is one step in a longer piece of work.

What you can ask it

"Which rows in the Pipeline table have Status Done and no owner?" The agent reads the base's tables to get the table id, then lists rows with a filter formula such as AND({Status}='Done', {Owner}=''), returning only the columns you name. Long results follow Airtable's pagination up to the row cap and say when more remain.

"Add these 40 prospects, and update the ones whose email already exists." The agent upserts the records matched on the Email field. Airtable takes 10 rows per request, so the connector splits the batch and spaces the requests, up to 200 rows per call, and reports how many were created and how many updated.

"Comment on the Acme row that the follow-up went out Friday." The agent finds the record, then posts the comment on it. A file such as a quote PDF can be attached to an attachment column the same way.

Airtable pairs with Notion and Google Sheets when the same data lives in more than one place, and with HubSpot when a row should become a contact.

What the Airtable connector can't do

• It does not build interfaces or automations, which Airtable's own server does.
• Nothing in the schema can be deleted, and a field's type cannot be changed through the API. A select option created by mistake can only be removed in Airtable's interface.
• Deleting a base is impossible through the API.
• Typecast is off by default. It lets a data write create missing select options, so writing "Signed" into a select that only knows "Signee" fails instead of adding a duplicate option. Turn it on only when you want that.
• An update is a PATCH. Sending a full replace clears every column not sent, so it is off unless asked for.
• Airtable allows 5 requests per second per base. If it cuts a batch off, the response says how many rows were written and the rest can be retried 30 seconds later.
• The synced-table tool pushes CSV into a table set up as an Airtable API sync source and replaces its content on each push. It is not a general import.
• The token sees only the bases it was granted. A base added later needs to be granted to the token too.