Connect ChatGPT to your database

Two routes that work today

Both need a URL OpenAI’s servers can reach, so the setup starts the same way.

1. Connect your database to Faucet

npx @faucetdb/faucet serve

Open http://localhost:8080, create an admin account, and add your database with its host, port, database, username and password. Test connection checks it before you save. Setup guides: PostgreSQL, MySQL and MariaDB, SQL Server, Oracle, SQLite, Snowflake.

2. Create a read-only key

faucet role create --name chatgpt-readonly --verbs GET --service mydb
faucet key create --role chatgpt-readonly --label chatgpt   # prints the key once

Or create a Read only role and a key in the admin UI. Leave raw SQL off for any database ChatGPT can reach.

3. Expose Faucet over HTTPS

Expose Faucet, never the database port. Pick one:

• Reverse proxy: nginx or Caddy in front of Faucet with a TLS certificate.
• Cloudflare Tunnel: cloudflared tunnel --url http://localhost:8080 prints a temporary https:// URL for testing.
• ngrok: ngrok http 8080 prints a public https:// URL.

Check it from outside your network:

curl https://your-host/healthz
curl "https://your-host/api/v1/mydb/_table?limit=5" -H "X-API-Key: faucet_YOUR_KEY"

4a. Custom GPT Action

1. Create a GPT and add an Action.
2. Import the schema from https://your-host/openapi.json. Faucet generates this OpenAPI 3.1 spec from your live schema.
3. Set authentication to API key, auth type Custom, header name X-API-Key, and paste your key.

The GPT can then call endpoints such as GET /api/v1/mydb/_table/orders?filter=status = 'open'&limit=25. Filters work like a SQL WHERE clause and are always sent as bound parameters.

4b. OpenAI Responses API

If you build on the OpenAI API, pass Faucet as an MCP tool in the tools array:

{
  "type": "mcp",
  "server_label": "faucet",
  "server_url": "https://your-host/mcp",
  "headers": { "X-API-Key": "faucet_YOUR_KEY" },
  "require_approval": "never"
}

The model can then call Faucet’s tools, such as faucet_list_tables, faucet_describe_table and faucet_query, limited by the key’s role. Drop require_approval if you want to approve calls yourself.

Before handing a key to ChatGPT, the admin UI’s API explorer can send requests as that key, so you see exactly what it can read.

Why not a ChatGPT connector?

ChatGPT connectors authenticate with OAuth. Faucet’s REST API and MCP endpoint use API keys and Faucet doesn’t offer OAuth yet, so a connector can’t sign in. The two routes above send the key as a header instead.