Say you want to pick a batch of people from an orders database: more than three refunds last month, average order value of 200 yuan or more. You roughly remember the table name, but whether the field is called refund_status or status, and whether the timestamp records the order or the refund, still requires checking the docs. Paste the table-creation statement into a chat box, and ChatGPT can also give you a SQL query; then you have to verify the table name yourself, adjust the dialect, and switch back to the database client to run it.
The trouble is not in “writing,” but in “after writing.”
AI2SQL is aimed at exactly this part. It reads the database you have connected, writes SQL using real table and column names, and sends it in that database’s dialect; you click Run on a read-only connection, and the returned rows appear on the page. The website FAQ puts this plainly: if you just paste a single schema into a chat box and ask a question, it is not very different from asking ChatGPT to write SQL; the difference appears when you need a live database: you do not have to tell it the table structure, it knows it on its own, and the same capability can also be handed to your own AI agent. (AI2SQL website[1])
The machine takes over the “run” step
The product’s main loop has two steps, and the same page also lines up Explain, Optimize, Fix, and Format, actions for working with existing statements. Generate: ask a question in natural language, and it outputs SQL in the selected database dialect, with real table and column names. Run: you click once, and only then does it execute against a live read-only connection, with results shown on the page. Before a statement reaches the database it first passes through sqlGuard, which classifies the statement and blocks INSERT, UPDATE, DELETE, DROP, other DDL, multi-statement input, and functions that read server files; the default is read-only. (AI2SQL website[1]) Only on the demo database does the free tier give 3 queries per day, as stated on both the website homepage[1] and the pricing page[2]; the MCP page[3] says 5 SQL requests per day, shared with the web app, so the two places do not agree. It is enough for practice; connecting your own database requires Pro or Team.
Image: source AI2sql official image[4], retrieved 2026-11-11, original interface retained.
One layer deeper is MCP. It lets assistants such as Claude, ChatGPT, or Cursor connect to your database: the hosted version exposes 8 tools, and the assistant chooses among them based on the user’s request; run_query accepts only a single SELECT/WITH, returns at most 1,000 rows, and times out at 10 seconds, while writes, DDL, and multiple statements are rejected on the server side. Authorization uses AI2SQL account login, with no need to paste an API key into the client; credentials are encrypted, and the assistant only sees the results. (AI2SQL MCP page[3])
There is a scope that must be made clear: MCP’s live schema browsing and run_query only support PostgreSQL, MySQL/MariaDB, and Microsoft SQL Server. For Oracle, Snowflake, BigQuery, SQLite, and MongoDB, it can write SQL, but live querying is not yet supported. (AI2SQL MCP page[3]) The connector list on the homepage is longer than this, and the two are not talking about the same thing: one is “can write this dialect,” the other is “can connect and run.”
The person who contacted free users one by one
Founder Mustafa Ergisi said in an Indie Hackers retrospective that he decided from the start to bootstrap rather than raise funding: slower growth in exchange for control and sustainable building, and in hindsight this was the right choice for him. (Indie Hackers retrospective[5]) The first 10 paying customers were the hardest; his approach was to contact free users one by one, offer extended trials, and collect feedback. He describes the Product Hunt and Hacker News launches as “a good initial boost”; he attributes truly sustained growth to SEO and word of mouth. (same source)
This attribution deserves a pause, but it is still the founder’s attribution, not a result calculated by channel. Launches are one-time; the retrospective does not give channel-level numbers on whether SEO and word of mouth can keep bringing in customers. A solo founder allocating weight this way looks more like day-to-day operations than the $100K ARR in the retrospective title. It should be noted that the quoted retrospective excerpt does not mark a publication date and cannot be treated as a November 2026 operating snapshot; he himself writes that $1K MRR took 15 months, and that growth is not linear. The $100K ARR in the title and the $32.7K ARR in the sale-page message are not in the same disclosure, how long apart they are is unclear, and no curve can be assembled.
Another mistake he admits was in pricing. At first there was only one tier at $7/month, which he himself thought was too cheap; after surveying users, he gradually raised the price and split it into tiers; every increase made him nervous, but he says it actually improved conversion. (same source) In his “if I could do it again” list, he writes: raise prices earlier, hire earlier, especially customer support—the two hardest things during expansion were, one, infrastructure to handle more queries, and two, responding to people in time; his approach was to automate as much as possible and build a thick knowledge base to reduce tickets. (same source)
Three tiers, from practice to letting agents touch data
The $9 Start tier gives 100 queries a month; the website’s own profile for it is occasional users such as “school, side projects, learning.” Pro at $19 connects “run” to your own PostgreSQL, MySQL, or SQL Server database: connect the database, run read-only queries, and get an API key for your own agent; the desktop app handles localhost and intranet reachability. Add $20 to go to Team; what you are buying is not more queries, but 5 people sharing one workspace, masking names, email addresses, phone numbers, and so on in results before handing them to an agent, and viewing activity logs for each agent in the team; it is more like a product sold to small teams preparing to let AI agents touch customer data.
Annual billing reduces the three tiers to $7, $14, and $29 per month, a 25% saving; there is also a one-time $29 Lifetime plan that includes 500 AI queries per month. All three tiers require a card to start a 7-day trial; you are charged $0 today, and canceling before the trial ends incurs no charge. (AI2SQL pricing[2])
SQLAI.ai can also connect to a database and run read-only queries, but by default it does not run automatically: after the user reviews the SQL, they decide whether to run it. It says it supports 36 database engines, with live connections available for PostgreSQL, MySQL, SQL Server, Oracle, and MongoDB, and features varying by engine; when not connected to a database, you can import a schema via DDL/JSON/CSV. (SQLAI.ai website[6]) The difference is not in the binary of “who presses the run key.” On AI2SQL’s side, Run is likewise initiated by the user against a read-only connection; it additionally gives agents the ability to call read-only queries, with a server-side read-only guard blocking writes. Neither cited website page provides a reproducible SQL accuracy benchmark, so there is no way to rank which is more accurate on that basis.
The selling point is written in “refusal,” the answer stuck on “coming back”
Revenue is viewed at a point in time. On the TrustMRR page dated 2026-11-14, the fields list all-time revenue $177,578, MRR $2,410, and 176 active subscriptions; the page says revenue is verified via Stripe API key. (TrustMRR listing page[7]) The numbers in the founder’s message on the same page are $2,723 MRR, equivalent to $32.7K ARR, $168K lifetime revenue, and 85% margin (definition not stated). Both sets of numbers are labeled MRR, but the page fields have an update time and Stripe verification, while the message has no date and does not state its definition; the two sides do not match and cannot be connected into one line. The cited page does not list model and infrastructure bills, so profit remains unclear.
There are several more statements in the message worth looking at: revenue over the last six months was about $3K/month, churn improved from 10.6% to 8.2%, and retention “has never been seriously addressed”; the sale is to focus on a new product. (TrustMRR listing page[7]) This is an asset with a for-sale page; whether the transaction was completed, the page does not say.
Putting these statements together with the product creates a bit of a mismatch. The strongest part of AI2SQL is actually “refusal”: sqlGuard blocks write operations before a statement reaches the database, run_query lets through only a single SELECT/WITH, and 1,000 rows and 10 seconds are hard limits—all of it keeps AI’s hands on the read-only tabletop. But its own retention, the founder candidly says, has not been seriously worked on.
For a tool that keeps risk outside the database door, the question it ultimately has to answer is not in the database: whether people will come back next week.
[1] AI SQL Query Generator for Your Own Schema | AI2SQL
[2] Pricing — Start $9, Pro $19, Team $39 with a 7-day free trial | AI2SQL
[3] Text-to-SQL MCP Server for Postgres, MySQL & SQL Server | AI2SQL
[5] From Zero to $100K ARR: My AI2SQL Journey
[6] Generate, optimize, validate and explain SQL with AI - SQLAI.ai