Airtable MCP Server
Exposes the full Airtable Web API including document attachment upload and download, runs on Cloudflare Workers with OAuth 2.1 authentication, and supports file staging via R2 for unlimited attachment sizes.
README
Airtable MCP Server
A remote Model Context Protocol server that exposes the full Airtable Web API to Claude — including document (attachment) upload and download, which the official Airtable connector does not support.
It runs on Cloudflare Workers, authenticates each user with their own Airtable account via OAuth 2.1 (PKCE-S256), and stages file uploads through Cloudflare R2 so files of any size can be attached to Airtable and then removed from Cloudflare automatically.
Works as a claude.ai custom connector, in Claude Desktop, and in Claude Code.
Why this exists
Uploading a file from your computer into an Airtable attachment field — and pulling files back out — without you having to think about hosting, URLs, or cleanup. You point Claude at a record; Claude uploads the document; the server handles staging, ingestion, and teardown.
How document upload works
Airtable ingests attachments from a publicly reachable URL, which it then re-hosts on its own storage. This server automates that safely, supporting files up to Airtable's 5 GB per-file limit:
- The file is staged in a private R2 bucket — small files inline as base64, large files streamed straight to R2 via a presigned PUT URL (bytes never pass through the Worker, so there's no ~100 MB request-body ceiling).
- The server hands Airtable a short-lived presigned GET URL for the staged object.
- Airtable fetches and re-hosts the file.
- The server polls the record until Airtable's re-hosted copy appears, then deletes the staged file from R2.
Presigned URLs are generated with an R2 S3 API token; the bucket itself stays private.
Architecture
| Piece | What it does |
|---|---|
McpAgent (Durable Object) |
Hosts the MCP server + tools, one instance per user session |
@cloudflare/workers-oauth-provider |
Makes the Worker an OAuth 2.1 server for MCP clients (claude.ai does dynamic client registration) |
| Airtable OAuth (upstream) | Each user signs in with their own Airtable account; tokens stored in KV, auto-refreshed (with rotation) |
| R2 bucket | Temporary staging for attachment bytes |
| Presigned R2 URLs | Client PUTs large files straight to R2; Airtable GETs the staged file to ingest — bytes bypass the Worker |
Transport: Streamable HTTP at /mcp (plus legacy /sse).
Tools
Records — list_records, get_record, create_records, update_records (PATCH + upsert), replace_records (PUT), delete_records (create/update/delete auto-batch in groups of 10)
Schema — list_bases, get_base_schema, create_base, create_table, update_table, create_field, update_field
Comments — list_comments, create_comment, update_comment, delete_comment
Webhooks — create_webhook, list_webhooks, delete_webhook, list_webhook_payloads, refresh_webhook, manage_webhook_notifications
Attachments — upload_attachment, create_attachment_upload_url, download_attachment
User — whoami
The server ships a detailed instructions block so Claude picks the right tool and follows the correct sequence (resolve IDs → read/write; stage → attach → cleanup) with minimal prompting.
Setup & deploy
Prerequisites
- Node.js 18+ and the Cloudflare Wrangler CLI (
npm i -g wrangleror usenpx), logged in (wrangler login). - R2 enabled on your Cloudflare account (Dashboard → R2 → enable). Required before deploy.
- An Airtable account.
1. Install
npm install
npm run cf-typegen # generates worker-configuration.d.ts
2. Create resources
npx wrangler kv namespace create airtable-mcp-server-OAUTH_KV # put the id in wrangler.jsonc
npx wrangler r2 bucket create airtable-mcp-files
3. First deploy (to get your Worker URL)
npm run deploy
Note the URL, e.g. https://airtable-mcp-server.<your-subdomain>.workers.dev.
4. Register an Airtable OAuth integration
Go to https://airtable.com/create/oauth → Register new OAuth integration.
- Name: anything (e.g. "My Airtable MCP").
- OAuth redirect URL:
https://airtable-mcp-server.<your-subdomain>.workers.dev/callback - Scopes (must match the server):
data.records:read,data.records:write,data.recordComments:read,data.recordComments:write,schema.bases:read,schema.bases:write,webhook:manage,user.email:read. - Copy the Client ID, then generate and copy the Client secret (shown once).
5. Create an R2 S3 API token
Cloudflare dashboard → R2 → Manage API Tokens → Create API Token (Object Read & Write). Copy the Access Key ID and Secret Access Key.
6. Set secrets
npx wrangler secret put AIRTABLE_CLIENT_ID
npx wrangler secret put AIRTABLE_CLIENT_SECRET
npx wrangler secret put R2_ACCESS_KEY_ID
npx wrangler secret put R2_SECRET_ACCESS_KEY
R2_ACCOUNT_ID is already set as a var in wrangler.jsonc.
7. Redeploy
npm run deploy
Connecting
- claude.ai → Settings → Connectors → Add custom connector → paste
https://<your-worker>/mcp. Sign in with Airtable when prompted. - Claude Desktop → Settings → Connectors → add the same
/mcpURL. - Claude Code →
claude mcp add --transport http airtable https://<your-worker>/mcp
Large files: truly unlimited uploads work best from Claude Code, which can stream bytes to the signed upload URL. From claude.ai/Desktop, files pass through the conversation and are bounded by context size.
Local development
cp .dev.vars.example .dev.vars # fill in secrets
npm run dev
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。