<!-- Source: https://docs.snoutdata.com/agent/mcp -->

# Agent access (MCP)

SnoutData can let an external agent, such as Claude Code, query your databases **through the
app** over the Model Context Protocol (MCP). The agent talks to a small local server that
SnoutData runs; your connection passwords and keys never leave this computer and never touch
the agent or a repo.

![The Agent access settings: local server, per-connection access, and the MCP config](https://docs.snoutdata.com/img/screenshots/agent-mcp.png)

## Turn it on

Open **Settings, Agent access** and enable **Allow agent access**. SnoutData starts a local
server (for example `http://127.0.0.1:7311/mcp`) protected by a bearer token.

## Choose what the agent can reach

Access is **off by default** for every connection. Tick only the connections you want an
agent to see; everything else stays private. For each ticked connection:

- Queries run **read-only** by default (`SELECT`/`WITH`/`SHOW`/`DESCRIBE`/`EXPLAIN`).
- Writes and DDL are refused unless you also enable **Allow writes** for that connection.

## Connect your agent

The settings panel gives you a standard MCP config block to add to your client (Claude Code,
Codex, Cline, and similar), and a one-line command for specific clients, for example:

```bash
claude mcp add-json snoutdata --scope user '{"type":"http","url":"http://127.0.0.1:7311/mcp","headers":{"Authorization":"Bearer <your-token>"}}'
```

The token authorizes the agent; keep it private. Use **Regenerate token** to revoke the old
one.

## Auditing

Queries the agent runs are recorded in your query history, the same as queries you run
yourself, so there is always a record of what an agent did.
