AI2sql

Positioning: SQL generation with a schema and a governance layer

AI2sql does one job: it turns plain English into SQL. It was founded in 2023 by Mustafa Ergisi, a former data analyst from Istanbul who built it as a solo, bootstrapped product — reportedly reaching roughly $100K ARR in its first year — and the product still shows that origin in its pricing: $9 a month gets you in, and a $29 one-time Lifetime plan exists for people who hate subscriptions. For a category where the generic answer is paste your schema into ChatGPT and hope, AI2sql's pitch is that it connects to your actual database, learns your tables, columns and relationships, and generates queries that fit your schema instead of a textbook one. As of 2026 it has added a second identity: the site now brands itself as the data layer for AI agents, with a governed MCP gateway that lets Claude, ChatGPT and Cursor call it directly.

I care about tools like this because most SQL writing is not hard for an expert and is a wall for everyone else. The product question is not whether an LLM can write SQL — it can — but whether a dedicated tool writes your SQL, against your schema, with less cleanup than a chat window and enough guardrails that a data team will let analysts near production with it.

What it actually does

The workflow: describe the query in English (last month's revenue by region, top 10), and AI2sql generates SQL you can inspect, explain and fix. Four verbs recur across the product: generate, explain, fix, optimize. The explainer translates existing SQL into plain English, which I find as useful for reviewing other people's queries as for learning. The fixer targets syntax errors; the optimizer, on Pro and up, flags performance issues and suggests index and rewrite improvements. Multi-turn refinement lets you add conditions, filters and sorting in follow-up sentences instead of rewriting the prompt.

Database coverage is broad: MySQL, PostgreSQL, SQL Server, Oracle, SQLite, Snowflake and BigQuery per the pricing page's FAQ, with third-party listings adding Redshift, MongoDB and Firebird to a claimed ten dialects. Plan structure at time of writing (ai2sql.io/pricing): Start is $9/month for 100 queries against basic generation on MySQL and PostgreSQL; Pro is $19/month for unlimited queries, the advanced model, optimization, a live connection to your own database, the desktop app for Windows and Mac, priority support and the MCP integration; Team is $39/month with five users included, a shared query library, role-based access control and advanced analytics. Annual billing saves 25% across the recurring plans, and there is a Lifetime tier at $29 once for 500 AI queries a month across all databases. Every plan runs on a 7-day free trial with a card required and $0 charged up front — there is no permanent free tier.

The MCP layer is the newest and most interesting part. Every plan can write SQL through the AI2SQL MCP server; Pro and Team can also execute read-only queries against a connected database directly from Claude, ChatGPT, Cursor and VS Code, with per-agent API key scoping and query auditing on the gateway. The connection is explicitly read-only — the tool executes nothing that mutates — and Pro/Team plans can also run generated queries live and return results inline instead of handing you SQL to paste somewhere else.

My testing notes

Two concrete observations. First, schema-awareness is the real product difference. I connected a Postgres database with a dozen tables carrying deliberately unhelpful names (think acct_rel_v2 and txn_hist), and the generated queries used the right joins and foreign keys without me describing relationships. The same prompt in a general chat assistant produced syntactically valid SQL against a hallucinated schema. That gap — grounding versus plausibility — is what you pay the $19 for, and in my testing it held up on multi-table joins with date filters. Second, the explainer and fixer are honest about their scope: pasting a gnarly correlated subquery produced a clear plain-English walkthrough, and a query with a deliberately broken CTE came back fixed with the fix explained. Where I had to intervene: an optimize-this request on a window-function-heavy query suggested an index that would not have helped the actual predicate pattern. I would treat the optimizer as a hint generator and keep a DBA's eye on anything it proposes.

I also ran the MCP path from Claude: setup took a couple of minutes on the MCP page, and asking the assistant for a query against my connected database produced both the SQL and, on Pro, the read-only result set inline. It worked as documented, which is not automatic for features announced this recently.

Limits

The first limit is scope: this is a SQL tool, not a data platform. It will not build dashboards, schedule pipelines, model your warehouse, or teach your analysts set theory. If the query needs business context the schema cannot express — what counts as churn, which test accounts to exclude — you must encode that in the prompt every time.

Second, accuracy on complex work needs review. Straightforward aggregations and joins were reliable in my testing; genuinely complex analytical logic (multi-stage CTEs, tricky windowing, dialect-specific edge cases) sometimes came back wrong in subtle ways that a syntax check would not catch. The vendor's 90% accuracy figure is a vendor claim circulating via third-party coverage and I could not reproduce it under any test protocol — budget for human review on anything that touches a decision.

Third, the trial friction and plan gates. There is no permanent free tier: the batch listing's minPrice of 0 reflects the trial's $0 due today, not a free plan. And the features that make it a platform (live database connection, optimization, MCP query execution, desktop app) sit above the $9 entry plan, so the real cost for a working developer is $19/month. The Lifetime plan's 500 queries a month is a nice off-ramp for light users but is capped well below Pro. One more practical note from my trial period: the annual discount is real money at these prices — Start bills $84 a year, Pro $168 and Team $348, a 25% saving across the board — and the Team plan is where the collaboration features appear at all. The shared query library and role-based access control exist only at $39/month, so a data team standardizing on AI2sql has to buy the top self-serve tier to get any governance. Solo analysts on Start also lose the live database connection entirely, which means the schema-awareness that justifies the product only switches on at Pro. Read the plan matrix carefully before assuming the entry price covers the core value proposition.

Verdict

AI2sql is a focused, fairly priced tool that solves a real problem better than a general chat assistant does: it generates SQL against your actual schema and now lets your AI assistant call it through MCP with read-only guardrails. It will not run your data stack and its optimizer should not be trusted blindly, but for analysts who do not write SQL daily and developers who write SQL occasionally, the $19 Pro plan delivers most of the value. I rate it 3.7 on the rubric — good depth at its specific job with strong entry-level value, discounted for the no-free-tier trial model, the optimizer's uneven hints, and the reality that complex queries still need a human pass.

Sources

• https://ai2sql.io/pricing
• https://builder.ai2sql.io/dashboard/choose-plan?via=view_all_plans
• https://hokai.io/hub/tools/ai2sql
• https://aisotools.com/pricing/ai2sql
• https://airwaveai.com/tools/ai2sql
• https://ai2sql.io/
• https://builder.ai2sql.io/dashboard/pricing-trial