SQAI — deterministic, read-only structured data MCP server

SQAI — Structured Query AI

SQAI is the deterministic, read-only structured-data tool for AI agents, with governed execution and replayable provenance.

Apache-2.0 · TypeScript + Python · built on Algenta as the underlying substrate · a Thyn product

The model proposes meaning. SQAI controls execution: typed intent in, policy-checked deterministic execution, replayable results with full provenance out. No raw SQL, no eval, no write path.

Architecture

Quickstart (AI SDK)

import { generateText } from "ai";
import { createSQAI } from "@thyn-ai/sqai-ai-sdk";

const sqai = createSQAI({ sources: [{ data: "./data/sales.csv", name: "sales" }] });

const { text } = await generateText({
  model: "openai/gpt-5-mini",
  tools: sqai.tools(),
  prompt: "Which region had the highest total revenue?",
});

No server. No daemon for the query plane. The computation plane provisions a signed managed runtime transparently on first use.

<!-- mcp-name: io.github.thyn-ai/sqai -->

MCP server

The same three governed tools are available to any MCP host (Claude Desktop, Cursor, …) as a stdio server — descriptions, Zod schemas, and execute functions are read straight off the AI-SDK tool objects, never forked:

npx @thyn-ai/sqai-mcp

{
  "mcpServers": {
    "sqai": {
      "command": "npx",
      "args": ["-y", "@thyn-ai/sqai-mcp"],
      "env": {
        "SQAI_SOURCES": "[{\"data\": \"./data/sales.csv\", \"name\": \"sales\"}]"
      }
    }
  }
}

Zero environment is a valid configuration: initialize and tools/list answer with no credentials (sources connect lazily on the first tool call). SQAIAPIKEY and SQAIDEPLOYMENTURL keep their usual semantics from @thyn-ai/sqai.

All three are annotated readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false. The computation plane is free on 1 machine with a one-time device registration (sqai login) — licensing moves a signed token, never your data; without it, computation calls return a structured licensing error by design and queries keep working.

What works today

• Query plane (connect / resolve / query / verify / ask) — in-process, both languages
• Full SQAI computation plane behind capability-contract validation
• Policy allow-lists (sources, fields, functions) the model can never override
• Cross-language conformance: identical values, planhash, intentsignature
• Model-context truncation with opaque result handles

Repository layout

Shortened. The full README is on GitHub.