🗄️ a2db

Agent-to-Database

Give AI agents safe, read-only access to your databases. One call, multiple queries, clean results.

5 databases · batch queries · pre-configured connections · SQLGlot read-only

Quick Start · MCP Tools · Security · Comparison · Setup

Agent: "Show me active users and their recent orders"
  ↓
a2db execute → 2 queries, 1 call, structured results
  ↓
Agent: "Got it — 847 active users, avg order $42.50"

Why a2db?

Most database MCP servers make you run one query at a time, repeat connection details on every call, and return results double-encoded inside JSON strings. a2db fixes all of that:

• Pre-configured connections — define databases in .mcp.json with --register, agent queries immediately
• Batch queries — run multiple named queries in a single tool call
• Default connection — set connection once, use it across all queries in a batch
• Clean output — structured JSON envelope with compact TSV data and per-query timing (see why TSV?)
• Read-only enforced — SQLGlot AST parsing blocks all write operations
• All drivers bundled — pip install a2db and you're done
• Secrets stay in env — ${DB_PASSWORD} in DSNs, expanded only at connection time

Supported Databases

Quick Start

pip install a2db

As an MCP Server (recommended)

Claude Code (with pre-configured connection):

claude mcp add -s user a2db -- a2db-mcp \
  --register myapp/prod/main 'postgresql://user:${DB_PASSWORD}@host/mydb'

Claude Code (minimal — agent calls login on demand):

claude mcp add -s user a2db -- a2db-mcp

Claude Desktop / Cursor / any MCP client (.mcp.json):

{
  "mcpServers": {
    "a2db": {
      "command": "uvx",
      "args": [
        "a2db-mcp",
        "--register", "myapp/prod/main", "postgresql://user:${DB_PASSWORD}@host/mydb"
      ],
      "env": {
        "DB_PASSWORD": "your-password-here"
      }
    }
  }
}

Multiple databases:

{
  "args": [
    "a2db-mcp",
    "--register", "myapp/prod/main", "postgresql://user:${DB_PASSWORD}@host/maindb",
    "--register", "myapp/prod/analytics", "postgresql://user:${DB_PASSWORD}@host/analytics"
  ]
}

--register pre-registers connections at server startup — the agent can query immediately. Passwords use ${ENV_VAR} syntax and are expanded at connection time, never stored in plaintext.

As a CLI

# Save a connection (validates immediately)
a2db login -p myapp -e prod -d main 'postgresql://user:${DB_PASSWORD}@localhost/mydb'

# Query
a2db query -p myapp -e prod -d main "SELECT * FROM users LIMIT 10"

# JSON output
a2db query -p myapp -e prod -d main -f json "SELECT * FROM users LIMIT 10"

# Explore schema
a2db schema -p myapp -e prod -d main tables
a2db schema -p myapp -e prod -d main columns -t users

# List / remove connections
a2db connections
a2db logout -p myapp -e prod -d main

MCP Tools

execute — the core tool

Named dict with default connection (preferred):

{
  "connection": {"project": "myapp", "env": "prod", "db": "main"},
  "queries": {
    "active_users": {"sql": "SELECT id, name FROM users WHERE active = true"},
    "recent_orders": {"sql": "SELECT id, total FROM orders ORDER BY created_at DESC LIMIT 5"}
  }
}

List format (auto-named q1, q2, ...):

{
  "connection": {"project": "myapp", "env": "prod", "db": "main"},
  "queries": [
    {"sql": "SELECT COUNT(*) AS cnt FROM users"},
    {"sql": "SELECT AVG(total) AS avg_order FROM orders"}
  ]
}

Response (TSV format — default):

{
  "active_users": {
    "data": "id\tname\n1\tAlice\n2\tBob\n3\tCharlie",
    "rows": 3,
    "truncated": false,
    "time_ms": 12
  },
  "recent_orders": {
    "data": "id\ttotal\n501\t129.00\n500\t49.99",
    "rows": 2,
    "truncated": false,
    "time_ms": 8
  }
}

No ::text casts needed — integers, floats, timestamps, arrays, NULLs all work natively.

Error context

When a query fails with a column error, a2db enriches the message:

column "nme" does not exist
Did you mean: name?
Available columns: id (integer), name (text), email (text), active (integer)

Why TSV?

LLM context windows are expensive. JSON row data is verbose — every row repeats every column name, adds braces, commas, and quotes. TSV is a flat grid: one header row, then just values separated by tabs.

For a 100-row, 5-column result set, TSV typically uses 40-60% fewer tokens than JSON row format. The structured JSON envelope still gives you metadata (row count, truncation status) — only the row payload is TSV.

Set format="json" if you need full structured output with column names on every row.

Security

Read-Only Enforcement

Every query is parsed by SQLGlot before execution:

• Blocked: INSERT, UPDATE, DELETE, DROP, TRUNCATE, ALTER, CREATE, GRANT, REVOKE
• Bypass-resistant: multi-statement attacks and comment-wrapped writes are caught at the AST level, not just keyword matching
• Allowed: SELECT, UNION, EXPLAIN, SHOW, DESCRIBE, PRAGMA

This is defense-in-depth — you should also use a read-only database user, but a2db won't let writes through even if the user has write permissions.

Write support is implemented in the core but not yet exposed via MCP. Planned: per-connection write permissions, explicitly enabled by the human operator — not the agent. See TODO.md.

Credential Storage

Connections are saved in ~/.config/a2db/connections/ as TOML files.

• ${DB_PASSWORD} syntax — environment variable references are stored literally and expanded only at connection time. Secrets stay in your environment, not on disk.
• No secrets in list output — list_connections shows project/env/db and database type, never DSNs or passwords
• Connection files are local to your machine and outside any repository

Deployment Scope

a2db currently runs as a local stdio MCP server. It inherits environment variables from the process that launches it (your shell, Claude Code, Docker). This is the standard model for local MCP servers — the same approach used by DBHub, Google Toolbox, and others.

Planned: remote HTTP transport with OAuth 2.1 per the MCP spec. For now, if running in Docker, inject secrets via environment variables at container runtime.

Comparison

When to use what:

• a2db — multi-DB batch queries with clean output, agent-first design, fast setup
• DBHub — custom tools via TOML config, web workbench UI
• Google Toolbox — GCP ecosystem, IAM integration, 40+ sources
• PGMCP — natural-language-to-SQL for PostgreSQL (requires OpenAI key)
• Supabase MCP — full Supabase platform management (edge functions, branching, storage)

Setup by Environment

Local (macOS / Linux)