内容
# Postgres MCP Server
[](#2-add-postgres-mcp-to-your-coding-agent)
[](#requirements)
[](./CHANGELOG.md)
**Connect your coding agent to PostgreSQL.** `postgres-mcp` lets you generate
queries and run analytics, design database schemas, diagnose performance of the
server and queries, securely manage connections, and import data. It is
compatible with any MCP client — GitHub Copilot, Claude Code, Codex, Open Code,
Cursor, VS Code, and more.
- **npm:** https://www.npmjs.com/package/@microsoft/postgres-mcp
- **No install required** — run it straight from `npx`.
## Quick start
### 1. Save a connection profile
Create a named profile once — the password is stored in your OS keyring, never in
a config file:
```sh
# add a profile from a libpq URI or key=value string
npx -y @microsoft/postgres-mcp connection add local \
"postgresql://postgres@localhost:5432/postgres"
# store its password securely (hidden prompt)
npx -y @microsoft/postgres-mcp connection set-password local
# check it
npx -y @microsoft/postgres-mcp connection list
```
> Profiles created by `connection add` permit write tools unless
> `access_mode: ro` is set. For autonomous agents that should not write, use an
> explicit read-only profile **and** a read-only database role — see
> [Read-only profiles](./USAGE.md#read-only-profiles-for-autonomous-agents).
### 2. Add postgres-mcp to your coding agent
The server speaks MCP over stdio. Point your client at
`npx @microsoft/postgres-mcp run` — no credentials go in the client config; the
coding agent discovers and connects to your saved profile at runtime.
**GitHub Copilot CLI, Claude Code, Cursor & Claude Desktop** — clients that use the `mcpServers` format (`~/.copilot/mcp-config.json`, `mcp.json`, `.cursor/mcp.json`, `claude_desktop_config.json`, …):
```jsonc
{
"mcpServers": {
"postgres": {
"command": "npx",
"args": ["-y", "@microsoft/postgres-mcp", "run"]
}
}
}
```
> Running headless or in CI, without a keyring? The server can also take a
> database connection string from the environment — see
> [Connecting without a profile](./USAGE.md#connecting-without-a-profile) in the
> guide.
### 3. Ask your coding agent
> "List the tables in my PostgreSQL database."
> "Show me the 10 most recent orders from my PostgreSQL database."
> "What's slowing down my PostgreSQL server right now?"
The agent calls the server's tools (`postgres_mcp_connect`,
`postgres_mcp_query`, `postgres_mcp_db_context`, …) for you.
---
## What it can do
| Area | Tools |
|------|-------|
| **Connections** | list / add / remove profiles, connect, disconnect, list databases |
| **Query** | read‑only SQL (`postgres_mcp_query`), DDL/DML (`postgres_mcp_modify`) |
| **Schema** | fetch `CREATE` scripts for tables, indexes, functions, sequences… |
| **Data** | describe a CSV, bulk‑load a CSV via `COPY` |
| **Diagnostics** | probe server capabilities, collect performance metric groups |
See the [usage guide](./USAGE.md) for tools, configuration, authentication
(including Microsoft Entra ID), TLS, and the security model.
---
## Security in one minute
- **The server doesn't decide what's allowed — your database does.** It just
passes requests along to PostgreSQL and runs them as the database user in
your connection profile. It can't tell the difference between a request you
meant to make and one the AI invented or was tricked into making, so
anything that database user can do, the AI can do. What actually protects
you is the permissions you give that user, the approval prompts in your MCP
client, and the model you choose to run. See
[Security model](./USAGE.md#security-model) for who's responsible for what,
plus a checklist for locking things down.
- **`postgres_mcp_query` is read‑only.** Omitted profile `access_mode` permits write
tools; set `access_mode: ro` and use a read-only database role unless writes
are intentionally delegated.
- **Microsoft Entra ID (AAD)** is selected automatically for Azure profiles
without a stored keyring password. Store a profile password to use password
authentication instead.
- **Local file reads** (`postgres_mcp_bulk_load_csv`, `postgres_mcp_describe_csv`) are limited
to the MCP server's startup working directory by default and paths allowed
with `allow-access-to-path <path>`. A file allows only itself; a directory
allows recursive access.
- **Secrets are scrubbed** from error messages before they leave the process.
See more info about security model in the [usage guide](./USAGE.md#security-model).
---
## Requirements
- **Node.js 22+** (to run via `npx`).
- Linux x64/arm64, macOS x64/arm64, or Windows x64. Windows ARM uses x64
emulation.
- A reachable **PostgreSQL** database — local, Docker, on‑premises, Azure
Database for PostgreSQL, Amazon RDS/Aurora, Google Cloud SQL/AlloyDB, or any
wire‑compatible server.
## Changelog
See [CHANGELOG.md](./CHANGELOG.md) for release notes.
## License
Licensed under the [MIT License](https://github.com/microsoft/postgres-mcp/blob/main/LICENSE) © Microsoft.
MCP 市场