Because it runs on Airtable's public API under the same identity you sign in with, a read-only collaborator gets a read-only connection and a workspace creator gets base creation, with no separate permission model to reason about. Which bases the assistant sees is managed on your third-party integrations page, so you can add or remove them without disconnecting.
Airtable's own hosted server at mcp.airtable.com/mcp, packaged here as an installable plugin for Claude Code, Codex and Cursor, and as standalone skills for any agent that reads the Agent Skills format.
- Questions answered against your own bases — searching and analysing records without knowing table IDs or field names
- Records created and updated in the bases you can already write to
- New bases created, for workspace owners and creators
- Interface data read, including where your access is interface-only, and new interfaces and interface pages created
- Automations inspected and drafted; switching one on is still done in the Airtable UI
- Installation that wires the server for you: a plugin for Claude Code, Codex and Cursor, or the skills alone through the skills CLI for any other agent
An Airtable account, connected over OAuth, plus permission to use third-party integrations in your organisation — an enterprise admin may need to allowlist it first. Access mirrors your existing Airtable permissions exactly: read-only and commenter access stays read-only here, and creating a base needs workspace owner or creator rights. It is a remote HTTP server, so nothing runs locally. Calls count against your normal Airtable API rate limits, and record creation is capped at 10 records per request.
One command — claude mcp add --transport http airtable https://mcp.airtable.com/mcp