Airtable MCP 服务

Better Airtable MCP

A better way to use Airtable with AI agents! The big features:

• SQL query support (backed by DuckDB)
• Preview & approval flow for any record creation, modification, or deletion
• Schema changes too: create and rename tables and fields (and edit their descriptions), behind that same approval flow

How to set up:

1. Go to https://claude.ai/customize/connectors
2. Create a new Connector titled Airtable with this URL: https://better-airtable-mcp.hackclub.com/mcp
3. Set permissions to "Always Allow" - better-airtable-mcp has an approval flow built-in and will prompt you in a web interface before it creates, modifies, or deletes records, or changes your base's schema

Full README generated by AI. Also see SPEC.md:

Hosted MCP server for Airtable with fast DuckDB-backed reads and human-approved writes.

This repository is being implemented from SPEC.md. The current implementation includes:

• OAuth provider flow for MCP clients chained to Airtable OAuth
• Shared per-base DuckDB caching with continuous sync workers
• Restart-safe sync restoration and startup cleanup of stale DuckDB files
• MCP tools for list_bases, list_schema, query, sync, mutate, manage_schema, and check_operation
• Human approval flow with a bundled React/Vite approval SPA
• Go unit and integration tests plus frontend unit tests

Toolchain

The repo uses mise to pin Go and Node:

mise install
mise exec -- npm --prefix frontend test
mise exec -- npm --prefix frontend run build
mise exec -- go test ./...
mise exec -- go build ./...

Container build:

docker build -t better-airtable-mcp:dev .

Local Development

Start Postgres only:

docker compose up -d postgres

Or start the full stack from Docker Compose:

docker compose up --build app

Load the local environment file:

set -a
source .env
set +a

Run the full verification suite with one command:

mise run verify

Or run the full test/build workflow manually and start the server locally:

mise exec -- npm --prefix frontend test
mise exec -- npm --prefix frontend run build
mise exec -- go test ./...
mise exec -- go run ./cmd/server

Once the server is running locally, open http://localhost:8080/debug to run the built-in OAuth flow, mint a bearer token for the local server, inspect the live MCP tool catalog, and execute tool calls manually.

For local Airtable OAuth testing, register this callback URL in your Airtable OAuth app:

http://localhost:8080/oauth/airtable/callback

BASE_URL=http://localhost:8080 is fine for a same-machine browser-based smoke test. If you want to test from a hosted MCP client or anything not running against your local machine directly, switch BASE_URL to a public HTTPS tunnel URL and update the Airtable callback to match exactly.

When running the app inside Docker Compose, the app container overrides DATABASE_URL to use the internal Postgres hostname, while the checked-in .env.example keeps the host-machine URL for local go run development.

Environment

The server requires these environment variables:

• DATABASE_URL
• AIRTABLE_CLIENT_ID
• AIRTABLE_CLIENT_SECRET
• BASE_URL
• APP_ENCRYPTION_KEY (must be exactly 32 bytes)

Optional, with defaults:

• PORT (default 8080)
• DUCKDB_DATA_DIR (default /data/duckdb) - where per-base DuckDB cache files are stored
• SYNC_INTERVAL_SECONDS (default 60) and SYNC_TTL_MINUTES (default 10)
• APPROVAL_TTL_MINUTES (default 60) - how long a pending approval stays valid
• QUERY_DEFAULT_LIMIT (default 100) and QUERY_MAX_LIMIT (default 1000)
• SYNC_MAX_CONCURRENT (default 4) - how many base syncs may run at once; the rest queue for a slot

These match the defaults described in SPEC.md.