AI data platformBeta

Datalic

Safe, verified AI over your own databases.

Datalic gives a team trusted answers from its own business data without hiring a data team. It comes in two shapes: Studio, the AI database client people use every day, and a platform of an SDK, an MCP server and embeddable widgets for putting the same governed layer inside your own product.

Request access See the engineering

At a glance

Who it is for: Finance, accounting and operations teams who need trusted answers from their own data, and product teams putting AI in front of a customer database.
Stage: Beta; Our lead product. In use, taking design partners.
Built with: PythonFastAPIDuckDBPostgreSQLReactElectronMCP

The problem

Why this product exists.

Teams are pointing AI at production databases, and the AI sounds equally confident when it is right and when it is wrong.

A wrong number in a board pack costs more than no number at all.

Handing database credentials to a cloud AI service is something most finance and regulated teams simply cannot sign off.

What it does

The cloud cannot open your database

The platform is split in two. The cloud brain reads the question and writes the query — it has no database driver and no credential code compiled into it. The SDK runs inside your network, holds credentials in memory only, and executes every query locally.

Studio — the AI database client

Connect to any database, browse every object, write and run queries in a real editor, and view, sort and edit results in a fast grid, with an AI analyst docked beside you. Ships as web, desktop and tablet apps.

Ask and Agent, one pipeline

Ask answers a single question end to end. Agent runs a multi-turn tool-using loop for harder work. Both stream over the same event vocabulary, so what you see never depends on which mode you picked.

32 databases, one governed interface

PostgreSQL, MySQL, SQL Server, Oracle, Snowflake, BigQuery, Databricks, Redshift, Trino, MongoDB, DynamoDB, Elasticsearch, Neo4j, Redis, ClickHouse, InfluxDB and the vector stores — each a separate connector package over a shared core.

Safe by default access control

Every saved connection carries a read-only or read-write mode plus optional database and table filters. The defaults are the safe ones, and the same guards apply to chat, to the SDK and to the MCP server.

A semantic layer that is not a model

Metrics, synonyms and column statistics turn a raw schema into meaning, so the agent references "monthly recurring revenue" rather than guessing at a SUM. It is deterministic Python — no embeddings, no model calls, and it records which rule matched.

One MCP server for twenty databases

Claude, Cursor, VS Code or any MCP client gets one server instead of ten, with the same access guards and answer verification as the chat product.

Runs where you need it

The same codebase runs local on a laptop with DuckDB, self-hosted in your own Docker with Postgres and S3-compatible storage, or in a managed cloud. It is a runtime setting, not a different edition.

Why it is hard to copy

Most tools in this space are a text-to-SQL demo with a login page. The hard parts are the ones nobody films: verification, a semantic layer that is explainable, per-connection access control, and an architecture where the cloud half is physically incapable of reaching your data.

Engineering behind it

A cloud service that cannot open your database

We split the product along that line. The cloud brain reads the question, picks the model and writes the native query — and is built with no database driver and no credential-handling code in it at all. The SDK runs inside the customer's network, holds credentials in memory only, connects to the database directly, and calls the brain one stateless step at a time.

Answers that check themselves — or abstain

Results are verified against the source before they are shown, and every figure carries the query and the rows behind it. The same execution path — access guards, query, verification — is used by the chat product, the SDK and the MCP server, so no entry point is a weaker door.

Teaching a schema what the business means

We built a deterministic semantic layer — metrics, bidirectional synonyms and real column statistics — in plain Python with no embeddings and no model calls inside it. Metrics are declarative data, scoped per connection, and every resolution records which rule matched.

One MCP server instead of ten

One server that discovers every installed connector and exposes a single consistent tool set across all of them, running on the same governed runtime as the chat product — so access guards and answer verification are inherited rather than reimplemented.

One codebase, three deployment shapes

Storage became a runtime decision instead of a build-time fork: DuckDB and the filesystem locally, PostgreSQL and S3-compatible storage in the cloud, with schema-per-organisation routing and lazy bootstrap. The desktop app holds both engine URLs at once and routes each request by the active workspace.

Next product

Datalic Sheets

An AI analyst for the spreadsheets you already have.

→

Interested in Datalic?

Tell us about your setup and we will show you it running on data like yours — not on a demo dataset.

Start a conversation About the company