io.github.ugurcl/dbridge-mcp

The agent discovers the schema on its own (list_tables, describe_table), then writes and runs a SELECT for whatever the user asks. No hand-written endpoint per question.

Works with SQLite, PostgreSQL, and MySQL / MariaDB.

Why

A raw LLM cannot know what is inside your database, and web search cannot reach private data. dbridge gives the model a guarded door to that data: it can read and answer, but it cannot write, drop, or leak the whole table.

Why dbridge and not another database MCP?

There are many database MCP servers. dbridge is built around one idea — you should be able to point an AI at a real database without holding your breath — and that shows up as a combination the single-engine servers don't offer:

• One server, three engines. SQLite, PostgreSQL, and MySQL/MariaDB behind the same seven tools and the same config. Switch databases by changing the connection string, not the tooling.
• Read-only twice over. A SQL guard rejects anything but SELECT/WITH and every query runs inside a database-enforced READ ONLY transaction — so even a query that outsmarts the guard cannot write.
• Column-level privacy. Hide columns from the model entirely, or mask values (a***@site.com) while keeping them queryable. Most servers expose whatever the connection can see.
• Blast-radius controls. Row caps enforced over user-supplied LIMITs, per-query timeouts, EXPLAIN-based cost rejection for expensive queries, per-minute rate limits, and a capped connection pool.
• No native build step. SQLite uses Node's built-in node:sqlite, so npx -y dbridge-mcp works without a compiler toolchain.

If all you need is "run SQL from my agent", plenty of servers do that. dbridge is for pointing an agent at data you actually care about.

Tools

Resources

Requirements

Node.js 22.5+ (SQLite uses the built-in node:sqlite, no native build step).

Install

The published package ships a dbridge-mcp binary, so no clone or build step is needed to use it. Point it at a database with the connection argument:

npx -y dbridge-mcp demo.db                                  # SQLite (file path)
npx -y dbridge-mcp "postgresql://user:pass@host:5432/mydb"  # PostgreSQL
npx -y dbridge-mcp "mysql://user:pass@host:3306/mydb"       # MySQL / MariaDB

The database engine is chosen from the connection string: a file path is SQLite, postgres:// / postgresql:// is PostgreSQL, and mysql:// is MySQL/MariaDB.

Or install it once, globally:

npm install -g dbridge-mcp
dbridge-mcp "postgresql://user:pass@host:5432/mydb"

MCP clients start the server for you as a subprocess — see the client sections below.

Windows note: MCP clients cannot launch npx directly on Windows because it is a .cmd script. Wrap it with cmd /c — use "command": "cmd" and put "/c", "npx", "-y", "dbridge-mcp", "<connection>" in the args. The examples below use the direct form (macOS/Linux); on Windows add the cmd /c prefix.

Use it in Claude Desktop

Claude Desktop only supports a single global config. Add to claude_desktop_config.json:

{
  "mcpServers": {
    "dbridge": {
      "command": "npx",
      "args": ["-y", "dbridge-mcp", "postgresql://user:pass@host:5432/mydb"],
      "env": { "DBRIDGE_CONFIG": "/absolute/path/to/dbridge.config.json" }
    }
  }
}

For a local SQLite file, replace the connection string with an absolute path to the .db file. Restart Claude Desktop, then ask: "what were the 5 best-selling products last month?"

Use it in Claude Code

Add it to the current project with the CLI:

claude mcp add dbridge --scope project \
  -e DBRIDGE_CONFIG=/absolute/path/to/dbridge.config.json \
  -- npx -y dbridge-mcp "postgresql://user:pass@host:5432/mydb"

--scope project writes a shareable .mcp.json in the project root; use --scope user for a global server or omit it for a private per-project one. The .mcp.json looks like:

{
  "mcpServers": {
    "dbridge": {
      "command": "npx",
      "args": ["-y", "dbridge-mcp", "postgresql://user:pass@host:5432/mydb"],
      "env": { "DBRIDGE_CONFIG": "/absolute/path/to/dbridge.config.json" }
    }
  }
}

Run claude from that directory; approve the project server once, then check it with /mcp.

Use it in Cursor

Add the same mcpServers block to .cursor/mcp.json in the project root (or ~/.cursor/mcp.json for all projects), enable the server under Settings → MCP, then ask the Agent a question about your data.

Use it in Windsurf

Add the same mcpServers block to ~/.codeium/windsurf/mcp_config.json, then refresh the server list under Settings → Cascade → MCP.

Use it in OpenCode

Add to opencode.json (project root or ~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "dbridge": {
      "type": "local",
      "command": ["npx", "-y", "dbridge-mcp", "postgresql://user:pass@host:5432/mydb"],
      "enabled": true
    }
  }
}

To tune the safety guard, point the server at a config file with the environment block:

"environment": { "DBRIDGE_CONFIG": "/absolute/path/to/dbridge.config.json" }

Docker

Build the image and run the server over stdio:

docker build -t dbridge-mcp .
docker run --rm -i dbridge-mcp "postgresql://user:pass@host:5432/mydb"

Mount a config file and point DBRIDGE_CONFIG at it:

docker run --rm -i \
  -v "$PWD/dbridge.config.json:/config.json:ro" \
  -e DBRIDGE_CONFIG=/config.json \
  dbridge-mcp "postgresql://user:pass@host:5432/mydb"

Safety

• The connection is opened read-only. On PostgreSQL and MySQL every query also runs inside a READ ONLY transaction, so writes are rejected by the database itself even if a query slips past the guard.
• Only SELECT and WITH statements pass; writes, DDL, and data-modifying CTEs are rejected.
• A single statement per call; the row cap is enforced even when a query supplies its own larger LIMIT (default 1000 rows).
• Each PostgreSQL and MySQL query is bounded by a per-query timeout (statement_timeout / max_execution_time), so a runaway or expensive query cannot pin the database.
• System catalogs and credential tables (information_schema, pg_authid, sqlite_master, …) are not queryable; schema discovery goes through the tools.
• Restricted columns can be hidden entirely: the model cannot see them in the schema, query them, or receive them in results.
• Tables can be restricted with an allow-list or block-list; blocked tables are invisible and unqueryable.
• Columns can be masked instead of hidden: they stay visible but values come back partially redacted (e.g. a***@site.com).
• Expensive queries can be rejected up front by an EXPLAIN cost estimate, and callers can be rate-limited per minute.
• The connection pool size is capped, so dbridge cannot exhaust the database's connections.

Config

Every setting has three sources, in increasing precedence: a JSON file (DBRIDGE_CONFIG), environment variables, then CLI flags. So you can drop the JSON file entirely and set only what you need:

npx -y dbridge-mcp "postgresql://user:pass@host/db" --max-rows 200 --statement-timeout-ms 3000 --masked-columns email,iban
DBRIDGE_MAX_ROWS=200 DBRIDGE_REQUIRE_SSL=true npx -y dbridge-mcp "postgresql://user:pass@host/db"

Or keep everything in one place: point DBRIDGE_CONFIG at a JSON file (see dbridge.config.example.json for a full template). Every field is optional:

{
  "maxRows": 500,
  "hiddenColumns": ["ssn", "password_hash"],
  "maskedColumns": ["iban", { "column": "email", "strategy": "email" }],
  "allowedTables": ["products", "sales", "customers"],
  "blockedTables": ["employees", "audit_log"],
  "statementTimeoutMs": 5000,
  "maxCost": 100000,
  "rateLimitPerMin": 60,
  "maxPoolSize": 5,
  "connectionTimeoutMs": 10000,
  "requireSsl": true,
  "schemas": ["public", "reporting"],
  "auditLog": true
}

List values on the command line or in env vars are comma-separated (--allowed-tables products,sales).

Engine note: statementTimeoutMs, maxCost, maxPoolSize, connectionTimeoutMs, requireSsl, and schemas apply to the networked engines (PostgreSQL/MySQL). SQLite is a local file, so it ignores them; the row cap, column/table access control, and masking apply to every engine.

Column masking

maskedColumns keeps a column visible but redacts its values. Each entry is either a column name (defaults to the partial strategy) or an object { "column": ..., "strategy": ..., "keep": ... }:

Unlike hiddenColumns, a masked column can still be used in WHERE/GROUP BY, so use hiddenColumns for true secrets and maskedColumns for values that should be recognizable but not exposed.

Output formats

run_query and sample_table take an optional format argument: json (default, full result object), csv, or markdown. CSV and Markdown return a compact table prefixed with a short rows: N · Nms header — handy for fewer tokens and readable output. Combine with maxCellChars and maxResultBytes to keep large results in check.

Running against a production database

• Pass the connection string via the DBRIDGE_DB_PATH environment variable instead of the command line, so the password does not appear in the process list.
• Prefer a dedicated database role with read-only grants on only the tables you want exposed — that is the real security boundary; the guard is defense in depth.
• Set a conservative statementTimeoutMs, maxRows, and maxCost, and use allowedTables to expose only reporting tables.

Local development

To hack on dbridge itself, clone the repo and build from source:

npm install
npm run build
npm run seed        # creates demo.db (a small store: products, customers, sales)
node dist/index.js demo.db

Try it with the MCP Inspector

npm run inspect

Then call the tools from the Inspector UI. No LLM or API key needed.

Tests

npm test

Changelog

See CHANGELOG.md.

License

MIT