Airtable MCP Server

Spreadsheet-database hybrid meets AI — inspect schemas, query records and update bases conversationally.

Local serverstdioTypeScriptMIT 400

What is the Airtable MCP server?

Airtable bases grow the way gardens do — enthusiastically and without consultation — until "which table has the client contacts, and what do these fields mean?" is a genuine research question. The Airtable MCP server (domdomegg's implementation is the community standard) gives your assistant both the map and the hands: full schema inspection across bases and tables, plus record-level read, create, update, delete and search.

Schema-awareness is what separates this from generic spreadsheet chat. The assistant first learns your actual structure — tables, field types, select options, linked records — then works within it correctly: records land with valid select values, links point at real rows, dates arrive formatted. "Add Globex to the CRM as a qualified lead, linked to yesterday's intro call" touches three tables correctly because the assistant read the schema first.

Queries handle the reporting layer Airtable's own interface makes you build views for: "deals closing this month over $10k", "content pieces with no assigned writer", "which projects reference this client?" — answered directly, no view-building required. Schema evolution is covered too (create tables, add fields), handy when a base needs to grow mid-conversation.

Setup is one personal access token with scopes you choose — schema read plus record read/write is typical — and granting base access at token creation is your blast-radius control. The npx one-liner makes this one of the fastest data integrations to stand up: token to talking-to-your-base in under five minutes.

Keeping the blast radius small

  • Personal access tokens are pinned to the bases ticked at creation and can't be widened later, so mint a dedicated one instead of reusing an automation token.
  • Leave schema.bases:write off unless you want tables and fields created; read scopes plus data.records:write cover every CRM-shaped workflow.
  • Formula, rollup, lookup and autonumber fields are computed, so writes aimed at them are rejected outright rather than quietly ignored.
  • Airtable rate-limits each base to five requests a second; asking for the specific fields you need keeps big sweeps comfortably under it.

What you can do with it

CRM upkeep

Leads logged, stages updated and calls linked — by sentence rather than by form.

View-free reporting

Filtered questions answered directly without building Airtable views first.

Base archaeology

Inherited bases explained: what tables exist, how they link, what fields mean.

Available tools

ToolWhat it does
list_bases / list_tablesEnumerate accessible bases and their tables
describe_tableField-level schema: types, options, linked records
list_records / search_recordsQuery and text-search records with filters
create_record / update_recordsAdd and modify records with schema-valid values
delete_recordsRemove records
create_table / create_fieldEvolve the schema when a base needs to grow

How to install the Airtable MCP server

{
  "mcpServers": {
    "airtable": {
      "command": "npx",
      "args": ["-y", "airtable-mcp-server"],
      "env": { "AIRTABLE_API_KEY": "pat_your_token" }
    }
  }
}

Create a personal access token at airtable.com/create/tokens with schema.bases:read and data.records scopes, granting only the bases you need.

Configuration

An Airtable account and a personal access token scoped to the bases you want exposed. Node.js 18+.

VariableDescriptionRequired
AIRTABLE_API_KEYPersonal access token with schema read + record read/write scopes on chosen basese.g. pat...Yes

Example prompts to try

  • Describe the structure of my CRM base — tables, key fields, and how they link.
  • Which deals close this month with value over $10k? List owner and stage.
  • Add Globex as a new lead: qualified stage, source 'referral', linked to the Acme intro call.

Frequently asked questions

At token creation — Airtable PATs are scoped to specific bases and permissions. Grant only what the assistant needs; the server can't see beyond its token.