The Nothing

[The Nothing: three dots on black]

The Nothing

a.k.a. Mr. Decoupler: the fastest private way to connect AI to a database.

Sometimes less is more.

[The Nothing MCP server]

The idea

Most database MCP servers hand the AI a query tool. The AI writes SQL, rows pour into the context window, and every question costs another round trip, more tokens and more exposure.

The Nothing turns that around. It holds the database connection privately, on your side, and gives the AI one minimal, read-only tool that can't run SQL:

• read_view schema: tables, columns and row counts. Structure only, never records.
• read_view <name>: a view you approved ahead of time, row-capped. Optional.

The AI can't write queries, browse tables, or change anything. It understands the shape of your data without dragging the data itself into the conversation. That's the decoupling, and it's why we call it Mr. Decoupler.

With no database around, it does exactly what the name says: nothing.

Zero setup

Open a project and it connects on its own, checking in this order:

1. THE_NOTHING_DB, if you set it
2. DATABASE_URL in the environment
3. THE_NOTHING_DB or DATABASE_URL in the project's .env file
4. A SQLite, CSV, TSV or JSONL file in the project folder

Nothing found means nothing connected. Auto-detection only looks inside a project folder, never your home folder or a drive root. Turn it off with THE_NOTHING_AUTODETECT=off.

How it works

 Your database / dataset ──(private, read-only)──▶ The Nothing ──(read_view: schema + approved views)──▶ AI   Postgres · MySQL · SQL Server                       │   SQLite · CSV · TSV · JSONL                          └─ the AI can't write SQL, browse or change data

Supported sources

Use a read-only database account. The Nothing enforces read-only on its side too, but a read-only account means nothing could write even by mistake.

Optional: approve views

When you want the AI to see actual numbers, approve views in a JSON file and point THE_NOTHING_VIEWS at it (see examples/views.example.json):

{  "views": {    "revenue_by_month": {      "description": "Total revenue per month.",      "sql": "SELECT date_trunc('month', placed_at) AS month, SUM(amount) AS revenue FROM orders GROUP BY 1 ORDER BY 1"    },    "customers_by_region": {      "description": "Counts only. No names or emails.",      "sql": "SELECT region, COUNT(*) AS customers FROM customers GROUP BY region"    }  }}

Aggregates make the best views: they give the AI what it needs to reason without raw rows in the conversation.

Install

VS Code + GitHub Copilot

Install The Nothing from the VS Code Marketplace and open your project. That's it: it auto-detects the database. To connect to something else, run The Nothing: Set Private Data Source from the Command Palette; the connection goes into VS Code's encrypted secret storage, never settings.json.

Claude Desktop

Download the-nothing-2.0.0.mcpb from the latest release and double-click it. Claude Desktop asks for the data source once and stores it as a sensitive setting. (Claude Desktop has no project folder, so this is the one place you paste a connection string.)

Cursor, Windsurf, Claude Code and everything else

Node.js 18+ is the only requirement.

{  "mcpServers": {    "the-nothing": {      "command": "npx",      "args": ["-y", "the-nothing-mcp"]    }  }}

If your editor starts MCP servers in the project folder, that's all; it auto-detects. Otherwise add "env": { "THE_NOTHING_DB": "postgres://readonly:password@localhost:5432/shop" }.

VS Code without the extension: .vscode/mcp.json

{ "servers": { "the-nothing": { "type": "stdio", "command": "npx", "args": ["-y", "the-nothing-mcp"] } } }

Visual Studio 2022 (17.14+) / 2026: <SolutionDir>\.mcp.json

{ "servers": { "the-nothing": { "type": "stdio", "command": "cmd", "args": ["/c", "npx", "-y", "the-nothing-mcp"] } } }

OpenAI Codex CLI: ~/.codex/config.toml

[mcp_servers.the-nothing]command = "npx"args = ["-y", "the-nothing-mcp"]

Remote clients (Streamable HTTP)

ChatGPT, Grok, claude.ai custom connectors and hosted agent APIs reach MCP servers over HTTPS:

THE_NOTHING_DB=postgres://… THE_NOTHING_HTTP_TOKEN=change-me npx -y the-nothing-mcp --http --port 3000

Clients send Authorization: Bearer <token>. With a data source, web mode refuses to start on a public address without a token, unless you set THE_NOTHING_ALLOW_PUBLIC_DATA=1 because a proxy in front already authenticates.

Configuration

All optional.

What the AI sees

When less is more

• Private by design: the AI learns your data's shape, not its contents.
• Memory and context over live queries: curated views you build once and reuse, instead of the AI re-querying on every turn.
• Keep agents away from production data: the AI can't query, browse or write, by construction.
• Fast: one tool, small answers, fewer tokens and round trips.
• Prototype before the database exists: the seam works with nothing behind it; add a source later and nothing on the AI side changes.

Development

npm installnpm test               # tools, transports, SQLite, CSV, JSONL, auto-detection, credential redaction, web-mode safetynpm run test:pg        # plus a real, temporary Postgres servernpm run build:vscode   # bundles the server into the VS Code extensionnpm run build:mcpb     # builds the Claude Desktop extension

See PUBLISHING.md for releasing.

FAQ

Is it really private? The connection and credentials stay in the server process on your machine. The AI only sees what read_view returns: structure and row counts, plus views you approved. Anything it reads is sent to your AI provider as part of the conversation, so keep views to aggregates where you can.

Why not a query tool? Because a query tool lets the AI decide what data enters the context window. The Nothing makes that your decision.

Can it write to my database? No. Views are read-only, run inside read-only or rolled-back transactions, and you should connect with a read-only account anyway.

Does it still do nothing? With no database, yes, exactly as in 1.x.

License

MIT

mcp-name: io.github.Meador793/the-nothing

...