airtable-kit

A type-safe, fully-featured, and standards-centric Airtable API client for TypeScript and JavaScript.

Motivation

There are other type-safe airtable clients out there (see the Comparison section), but they all start from the premise of starting from typescript interfaces to define the shape of records in each table. This package takes a different approach: everything is derived directly from the Airtable Base's JSON schema, as directly provided by the Airtable API. We then provide some typescript generic magic so that we infer that e.g. a singleSelect field with options "To Do", "In Progress", and "Done" maps to the union type "To Do" | "In Progress" | "Done" in typescript.

This has several advantages:

• It is trivial to codegen the schema files from existing bases since you pretty much write the JSON directly to a file, instead of needing to translate it to typescript interfaces. This makes it easy to create/update base schemas.
• You get runtime validation and JSONschema (eg for API docs/MCP servers/etc) for free. Most of these other libraries assume you know at compile time which bases you are working with. This package allows you to fetch a base's schema just-in-time at runtime, and still have type-mapping and runtime validation.
• Since we preserve the full JSON schema, you can inspect and manipulate the base's schema at runtime, eg generate a Zod validator. This is useful for e.g. building MCP tools for AI agents, since now they can see the exact schema of the base they are working with.

Other Features

• Rename-Resistant: When someone modifies the name of a field in airtable, your code won't break because the client always sends field IDs to the API under the hood.
• Complete Type-Safety: Not just the basics. Eg get type safety on the valid options for single-select and multi-select fields, date formats, etc.
• More-sane read behavior: The raw API (and vanilla airtable.js) don't include "falsy" values when reading records. We fill in those missing values with sane defaults, eg the empty string for text-like fields, empty arrays for multi-value fields, or an explicit null for eg number and date fields.
• Typesafe Formula Builder: Build Airtable formulas programmatically with full type-safety and autocompletion, eg formulaToString(tableSchema, ["<=", { field: "Due Date" }, ["DATEADD", ["TODAY"], 7, "days"]]) yields {fld123xyz} <= DATEADD(TODAY(), 7, "days").
• Both Functional and OOP APIs: Use either the functional API that is a 1:1 mapping to the Airtable REST API, or the OOP-style client with methods like base.tables.myTable.create(records).
• Zero-Dependencies: No required dependencies! zod is optional if you want runtime validation (eg for MCP tools).
• Portable: Works in Node.js, Deno, and the browser, including in sandboxed environments, e.g. where process.env and os.homedir() are not available.
• Runtime Validation: Optionally, dynamically generate Zod validators from the schema to validate all data sent to Airtable at runtime.
• MCP Tools: Ready-to-use MCP tools for LLM integration.

Installation

This is available on npm:

pnpm add airtable-kit

Quick Start

Use Case 1: Statically Typed Client for a Known Base

Option A: Use the Web App (No Installation Required)

Visit https://nickcrews.github.io/airtable-kit/ to generate schema files directly in your browser. Just enter your API key, select a base, and copy/download the generated code.

Option B: Use the CLI

Run the codegen CLI to fetch and save all your base schemas to local files:

npx airtable-kit codegen all --api-key YOUR_API_KEY

This will fetch all the bases that the API key has access for, and create files under ./schemas/, one per base. (If you instead wanted to just generate a single base schema, see npx airtable-kit codegen base --help.) Here is an example generated schema file for a base called Project Tracker, saved as schemas/projectTracker.ts:

export default {
  "id": "appXXXXXXXXXXXXXX",
  "name": "projectTracker",
  "tables": [
    {
      "id": "tblXXXXXXXXXXXXXX",
      "name": "tasks",
      "primaryFieldId": "fldXXXXXXXXXXXXXX",
      "fields": [
        {
          "id": "fldXXXXXXXXXXXXXX",
          "name": "name",
          "type": "singleLineText",
        },
        {
          "id": "fldYYYYYYYYYYYYYY",
          "name": "status",
          "type": "singleSelect",
          "options": {
            "choices": [
              { "id": "selAAAAAAAAAAAAAA", "name": "To Do", "color": "red" },
              { "id": "selBBBBBBBBBBBBBB", "name": "In Progress", "color": "blue" },
              { "id": "selCCCCCCCCCCCCCC", "name": "Done", "color": "green" }
            ]
          }
        },
        {
          "id": "fldZZZZZZZZZZZZZZ",
          "name": "dueDate",
          "type": "date",
          "options": {
            "dateFormat": "iso",
            "timeFormat": "24hour",
            "timeZone": "client"
          }
        },
      ]
    }
  ]
} as const;

Note how this is almost a direct 1:1 correspondence with the raw JSON from the GET https://api.airtable.com/v0/meta/bases/{baseId}/tables endpoint. There are a few minor differences, such as table and field names are converted to camelCase for easier consumption in javascript/TypeScript, and we add the top-level id: "appXXXXXXXXXXXXXX" and name: "projectTracker" fields, which aren't present in the raw JSON.

If you dropped the as const you would have a plain javascript file if you are working with javascript. You can get the CLI to do this with the --format js flag.

Now, use the generated schema to create a type-safe Airtable client:

import * as atk from 'airtable-kit';
import projectTrackerSchema from './schemas/projectTracker.ts';

const client = atk.clients.makeBaseClient({
  baseSchema: projectTrackerSchema,
  fetcher: 'YOUR_API_KEY',
});
// if you tried `client.tables.nonExistentTable` you would get a type error here
client.tables.tasks.createRecords([
  {
    // You can use either the field IDs or names as keys
    fldXXXXXXXXXXXXXX: "Finish report",
    status: "In Progress",
    // This would give a type error because its not a valid select option
    // status: "in progress",
    // You can either provide a Date object or an ISO string
    dueDate: new Date("2024-12-31"), // or "2024-12-31" works too
    // This would give a type error because 42 is not a valid date
    // dueDate: 42
    // This would give a type error because dueBy is not a field
    // dueBy: "2024-12-31"
  },
]);
const tasks = await client.tables.tasks.listRecords({
  // You can build formulas programmatically with type-safety
  // All tasks due in the next 7 days that are not done.
  // This is compiled to the string:
  // AND(
  //   {fldZZZZZZZZZZZZZZ} >= TODAY(),
  //   {fldZZZZZZZZZZZZZZ} <= DATEADD(TODAY(), 7, "days"),
  //   OR(
  //     {fldYYYYYYYYYYYYYY} = "To Do",
  //     {fldYYYYYYYYYYYYYY} = "In Progress"
  //   )
  // )
  filterByFormula: [
    "AND",
    [">=", { field: "dueDate" }, ["TODAY"]],
    ["<=", { field: "dueDate" }, ["DATEADD", ["TODAY"], 7, "days"]],
    ["OR",
      ["=", { field: "status" }, "To Do"],
      ["=", { field: "status" }, "In Progress"],
    ],
  ],
});
console.log(tasks);

This also supports multi-base clients!

import * as atk from 'airtable-kit';
import projectTrackerSchema from './schemas/projectTracker.ts';
import otherBase from './schemas/otherBase.ts';

const client = atk.clients.makeOrgClient({
  bases: [projectTrackerSchema, otherBase],
  fetcher: 'YOUR_API_KEY',
});
client.bases.otherBase.tables.someTable.createRecords([ ... ]);
// `client.bases.<base name>` is an instance of BaseClient, like above.
// `client.bases.nonExistentBase` would be a type error

Use Case 2: Dynamic Client for Unknown Bases, eg an MCP Tool

import * as atk from 'airtable-kit';

const fetcher = 'YOUR_API_KEY';
const baseSchema = await atk.bases.getBaseSchema({ baseId: 'appXXXXXXXXXXXXXX', fetcher });
const client = atk.clients.makeBaseClient({ baseSchema, fetcher });
const listTool = atk.mcp.makeListTool(client);

// Now you can pass this off to an AI agent framework, eg Vercel's AI SDK (https://github.com/vercel/ai)
import { tool, ToolLoopAgent } from 'ai'
const agent = new ToolLoopAgent({
  model: 'openai/gpt-5-codex',
  system: 'Help the user find info in their Airtable base.',
  tools: {
    listRecords: tool({
      description: listTool.description,
      inputSchema: listTool.zodInputValidator,
      execute: listTool.execute,
    })
  },
});

Use Case 3: Use the functional API for simple uses

import atk from 'airtable-kit';

atk.fields.createField(
  {
    baseId: 'appZZZZZZZZZZZ',
    tableId: 'tblAAAAAAAAAAAAAA',
    field: {
      name: 'newField',
      type: 'singleLineText',
    },
    fetcher: 'YOUR_API_KEY',
  }
)

atk.records.updateRecords(
  {
    records: [
      {id: 'recXXXXXXXXXXXX', fields: { 'fldYYYYYYYYYYYYYY': 47 } },
    ],
    baseId: 'appZZZZZZZZZZZ',
    tableId: 'tblAAAAAAAAAAAAAA',
    fetcher: 'YOUR_API_KEY',
    // TODO: This is currently required to do value conversion when reading
    // and writing to the API, but ideally this would be optional, in which
    // case we would just pass through values as-is.
    // See the "more sane read behavior" feature above.
    fields: [{
      id: 'fldYYYYYYYYYYYYYY',
      name: 'priority',
      type: 'number',
    }]
  }
)

Comparison to Other Airtable Libraries

airtable-ts

I was using this before building this library, and it is a great library.

• Also type-safe
• Derives types from typescript interfaces instead of the raw base schema. This loses some type-safety, eg for supplying valid select options. This is how most typescript airtable libraries work. There is a separate airtable-ts-codegen package to generate these interfaces.
• Requires doing two fetches for all operations: one to get the tables schema so that it can map field types and names, and another to do the actual operation (eg insert records). This package, since it operates directly from the base schema, therefore only needs to do one fetch per operation after the initial schema fetch.
• You can only work with bases known at compile-time, you can't dynamically fetch a base schema at runtime and make a client for it.
• No zod schema generation for JSONschema generation.

airtable-mcp-server

• Single purpose MCP server.
• Does not expose an API for building your own MCP tools or agents, where this package does.
• Does support working with unknown bases at runtime, similar to this package. But this package allows for generating a set of MCP tools per base, whereas that package only provides a single insertRecords tool that works with any base. This means that we can provide a specific-to-that-base JSON schema for the LLM to work with, which (I think, but not tested) improves reliability.

License

MIT

Contributing

I don't want this project to feature-creep too much, so at first I will be hesitant on accepting new features. At least file an issue before starting work. But bug reports, documentation improvements, and performance improvements are always welcome!