factql-mcp

FactQL is the harness around the model.

Ask a question about your data. FactQL runs the query, checks the result before you see an answer, and declines instead of guessing when it can't verify.

Every answer ships with the query that ran, the rows it returned and the lineage behind it. One API call. Works with Postgres, MySQL, DuckDB, uploaded CSV and XLSX files, and PDFs with cited answers.

Built for developers putting AI near real data: AI apps, agents and internal tools that can't afford a confident wrong number.

This package is FactQL's MCP server: it brings those answers into Claude Desktop, Claude Code, Cursor, or any MCP client. Learn more at factql.ai.

Get a key

Sign up free at factql.ai/signup, connect your data, and create an API key. An ask-only key is enough for this server.

Install

uvx factql-mcp

or pip install factql-mcp.

Configure

Claude Desktop / Cursor (mcpServers)

{
  "mcpServers": {
    "factql": {
      "command": "uvx",
      "args": ["factql-mcp"],
      "env": { "FACTQL_API_KEY": "fl_live_..." }
    }
  }
}

Claude Code

claude mcp add factql -e FACTQL_API_KEY=fl_live_... -- uvx factql-mcp

Tools

• whoami() — the tenant, plan, and workspaces this key can query.
• ask(question, workspace_id?, locale?) — the answer, the SQL that ran, and up to 50 rows. A needs_clarification result means the question was too vague; ask the user and call again with the detail. A declined result means FactQL couldn't verify an answer and gave none rather than guess; the reason says why. locale is en-GB or en-US.

Each ask is stateless: the MCP client carries the conversation, so a follow-up should restate what it refers to.

Safety

FactQL only runs SELECT queries, under a statement timeout, and only against the tables your admin has enabled. The key's permissions decide what it can see; this server adds nothing on top.

Development

uv run --with pytest pytest -q

License

MIT