| --- |
| title: SafeSQL-MCP |
| emoji: 🛡️ |
| colorFrom: blue |
| colorTo: gray |
| sdk: docker |
| app_port: 7860 |
| pinned: false |
| license: mit |
| short_description: A text-to-SQL MCP server an AI agent can't break |
| --- |
| |
| # SafeSQL-MCP: try to break it |
|
|
| A text-to-SQL MCP server for Postgres. Every query an agent writes is parsed into a syntax |
| tree, checked against allowlists, cost-checked with `EXPLAIN` and run read-only under a |
| read-only role. On this page you can type any SQL, or fire one of the 132 attacks from the |
| CI suite, and see which layer stops it. |
|
|
| - **Playground:** the page above. It runs your SQL through the same guard as the MCP tools. |
| - **MCP endpoint:** `https://<this-space>.hf.space/mcp` (streamable HTTP). Connect Claude Code, |
| Claude Desktop or Cursor and ask questions about the Spider `world_1` database. |
| - **Source, evals and CI gate:** https://github.com/ChaitanyaKharche/SafeSQL-MCP |
|
|
| Data: [Spider](https://yale-lily.github.io/spider) (Yu et al., 2018; CC BY-SA 4.0), ported to |
| Postgres, plus a synthetic shop table with planted prompt injections. |
|
|
