AI SQL Agent MCP Server

A production-hardened Model Context Protocol (MCP) server that lets any MCP-compatible AI client (Claude, LangGraph agents, etc.) ask natural-language questions about a SQL database — safely.

Why this exists

Most public "LLM writes SQL, we run it" examples stop at a single unguarded step: generate SQL, execute it, return rows. That's fine for a demo and dangerous in production — a single hallucinated DROP TABLE or an unbounded query is a real incident, not a hypothetical.

This project treats the SQL-generation step as untrusted input and puts a real safety pipeline around it.

Related MCP server: MCP-DB-Server

Architecture

MCP Client (Claude / LangGraph / MCP Inspector) │ ▼ FastMCP Server (this repo) │ ┌────┼─────────────────────────────┐ ▼ ▼ ▼ Schema LLM SQL Generation Guardrail Layer Reader (schema-grounded (read-only enforcement, (real prompt, no guessing) forbidden-keyword scan, tables) statement-stacking block, │ row-limit injection) └──────────────┬──────────────────┘ ▼ Read-only SQLite connection │ ▼ Structured JSON result + trace log

Features

• Schema-grounded generation — the LLM only ever sees real table/column names pulled live from the database, never guesses.

• Guardrail layer — blocks any non-SELECT statement, blocks statement-stacking (; DROP TABLE), strips comments before keyword scanning, and caps result size with an injected LIMIT.

• Defense in depth — the guardrail check runs and the DB connection itself is opened read-only.

• Optional bearer-token auth on the MCP tool.

• Structured, correlated logging — every request gets a request ID, logged SQL, latency, and outcome (success / rejected / error).

• 28 automated tests covering guardrails, schema introspection, generation, and the full integration pipeline — including adversarial cases (SQL hidden in comments, oversized LIMITs, disguised keywords).

• Health check endpoint (/health) for uptime monitoring.

Why this is different from a typical FastMCP demo

FastMCP (Apache-2.0, PrefectHQ) is used here as the underlying server framework — it handles MCP protocol plumbing so this project can focus entirely on the parts that matter for a real data-access agent: safety, grounding, and observability. Everything in guardrails.py, schema_reader.py, sql_generator.py, and observability.py is original to this project.

Tech stack

Python 3.11+ · FastMCP (Apache-2.0) · Anthropic API · SQLite · pytest · Docker · GitHub Actions

Installation

git clone https://github.com/rajmyagentit-del/ai-sql-agent-mcp.git
cd ai-sql-agent-mcp
pip install -e ".[dev]"

Environment variables

Usage

python -m ai_sql_agent_mcp.server

Connect with the MCP Inspector to try it interactively:

npx @modelcontextprotocol/inspector python -m ai_sql_agent_mcp.server

Testing

pytest -v

API

Exposes two MCP tools:

• ask_database(question: str, api_token: str | None) -> dict — the main tool.

• describe_schema() -> dict — returns the live database schema.

Verified Live

This isn't just claimed — here's what's independently checkable right now:

• Live demo: ai-sql-agent-mcp.onrender.com/demo — try it yourself, no login required. (Free-tier hosting: the first request after a quiet period can take up to ~50s while the server wakes up — that's expected, not a bug.)

• Health check: ai-sql-agent-mcp.onrender.com/health — should return {"status":"ok"}.

• CI status: the badge at the top of this README is live and updates automatically on every commit — click it to see real, current test runs, not a static image.

• Real MCP protocol handshake, captured directly against the live deployment via curl (not a mock):

$ curl -s -i -X POST https://ai-sql-agent-mcp.onrender.com/mcp -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl-test","version":"1.0"}}}'

HTTP/2 200 content-type: text/event-stream

event: message data: {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-06-18", "capabilities":{"tools":{"listChanged":true}, ...}, "serverInfo":{"name":"ai-sql-agent","version":"3.4.7"}}}

Real example — successful query (captured live, not fabricated):

// Question: "Show me all customers"
{
  "ok": true,
  "sql": "SELECT * FROM customers LIMIT 500",
  "columns": ["id", "name", "country"],
  "rows": [
    {"id": 1, "name": "Ada Lovelace", "country": "UK"},
    {"id": 2, "name": "Grace Hopper", "country": "USA"},
    {"id": 3, "name": "Alan Turing", "country": "UK"}
  ]
}

Real example — a destructive request refused (captured live):

// Question: "Delete all customers"
{
  "ok": false,
  "error": "cannot_answer",
  "reason": "The request requires a DELETE operation, which is not allowed. Only SELECT statements are permitted."
}

This particular case was caught at the generation layer — the model followed its system instructions and declined rather than emitting a DELETE statement. That's one layer of defense. The second, independent layer — the guardrail rejecting a write statement even if the model disobeys and generates one anyway — is proven separately by the automated test suite (see test_rejects_delete_disguised_via_comment, test_rejects_update, etc. in tests/test_guardrails.py), which deliberately feeds disallowed SQL straight past generation to confirm the guardrail alone stops it.

Evaluation Results

Real, captured results from evaluation/run_eval.py run against the live deployment (not invented, reproducible by anyone by running the script themselves):

{
  "accuracy_score": "7/7",
  "accuracy_pct": 100.0,
  "safety_score": "5/5",
  "safety_pct": 100.0,
  "avg_latency_ms": 1270.4
}

• Accuracy (7/7): natural-language questions correctly answered, verified against known-correct facts about the seeded dataset (counts, filters, aggregates, max-value lookups) — not just "did it return something," but "was the answer actually right."

• Safety (5/5): every destructive/injection attempt was refused — either at the generation layer (the model declining to write disallowed SQL) or the guardrail layer (structural rejection of any write/DDL/stacked statement that did get generated). One case was additionally blocked by the hosting provider's own edge security layer before even reaching the app — a genuine extra defense layer, tracked separately in the harness rather than credited to this project's own code.

• Reproduce it yourself: python evaluation/run_eval.py --url https://ai-sql-agent-mcp.onrender.com

Deployment

See docs/DEPLOYMENT.md for the free-tier cloud deployment guide.

Security

• Read-only by construction at two independent layers (guardrail + DB connection mode).

• No write, DDL, or multi-statement SQL can ever reach the database.

• API token auth available; recommend enabling it for any public deployment.

• No secrets are logged; only the generated SQL text and metadata are logged.

Limitations

• Currently supports SQLite; Postgres/MySQL support would extend SchemaReader.

• Guardrails are structural (keyword/statement based), not a full SQL parser — sufficient for this threat model but not a substitute for DB-level permissions in a high-stakes production environment.

Roadmap

• Postgres/MySQL schema reader

• Query result caching

• Evaluation harness (NL→SQL accuracy benchmark) — see Evaluation Results above

• Rate limiting

License

MIT — see LICENSE. Built on FastMCP (Apache-2.0, PrefectHQ).

Author

Built by Raj Dantuluri as part of an AI/ML engineering portfolio.