Connect Claude to Airtable

List the bases the token can reach, each with its base id, name and permission level. Call this first: every other operation needs a base id, and a base id is not something an agent can guess from a base's name.

Describe every table in one base: table ids and names, each field's id, name and type (including select options and linked-record targets), and the views defined on the table. Call this after list_bases and before any record operation - it supplies the table name or id, the exact field names a filterByFormula or a write has to use, and the view names list_records accepts. Guessing a field name instead returns an empty page rather than an error.

Return records from one table, optionally narrowed by a view and sorted. This is the browse operation: use it when you want a page of a table as a person would see it. Use search_records instead when you have a formula that selects the rows you want. Call get_base_schema first so `view`, `fields` and `sort_field` name things that exist. Paginate by passing the `offset` from the previous response.

Return the records of one table that match an Airtable formula. Prefer this over list_records whenever you know the condition - {Status} = 'Active', AND({Owner} = '[email protected]', {Due} < TODAY()), SEARCH('acme', LOWER({Company})). Field names go in curly braces and must match get_base_schema exactly: a formula naming a field that does not exist returns an empty page, not an error, so check the schema before concluding there are no matches.

Fetch one record by its record id, with every field it has a value for. Airtable omits empty fields entirely, so a field missing from the response is empty rather than absent from the table - get_base_schema is what tells you which fields exist. Record ids come from list_records or search_records.

Create up to 10 records in one table. Each entry is an object of field name to value, matching the types get_base_schema reports - a linked-record field takes an array of record ids, a multiple-select takes an array of option names. Set typecast to let Airtable coerce loose values (a date string, a new select option) instead of rejecting them. Returns the created records with their new record ids.

Update up to 10 existing records by record id, changing only the fields you name and leaving the rest as they are. Each entry is {"id": "recXXXX", "fields": {…}}. Use this rather than delete-then-create: it preserves the record id, its created time and every field you did not mention. Call get_record first if you need to know what a field currently holds.

Update records that match on one or more key fields, and create the ones that do not exist yet. Supply records without ids plus merge_on naming the fields Airtable should match against - an external id, an email address. The response separates createdRecords from updatedRecords so you can tell what happened. Prefer update_records when you already hold record ids; this exists for syncing from a source that only knows its own keys.