Your assistant stops guessing and starts reading your schema.
Ask an assistant to write a query against a database it has never seen and it will invent the table names, the column names, and the values, confidently, because inventing is the only option it has. Taavik hands it the real catalog instead. The connector goes in once and every answer after that is built on what your last scan actually found.
The problem
A query that returns nothing looks exactly like a query that works.
There are three ways for it to be wrong, and they are not equally dangerous. Only the first one makes any noise.
It invents a table that does not exist
The database rejects the query and you fix it by hand. Annoying, and completely harmless, because a failure that announces itself is one you never act on by mistake.
It guesses a value that is written differently
It writes WHERE status = 'completed' and your data says COMPLETED. Or 'C'. The query runs, raises nothing, and hands back an empty result that looks exactly like a correct answer, so whoever asked walks away believing there are no completed orders. This is the expensive one. It does not fail, it lies.
It cannot find a relation nobody declared
Plenty of production databases carry no foreign keys at all, sometimes for performance, more often because that decision was made years ago by someone who has since left. An assistant reading the constraints sees a pile of unrelated tables and joins them the way it hopes they connect.
How it works
Three steps, and the last one is the only one you repeat.
The agent scans
If you already use Taavik for documentation this happens anyway, and the catalog refreshes with it, which means yesterday's rename is in there without anyone exporting anything.
You add the MCP connector once
Point your assistant at Taavik and sign in. That is the whole setup, and no credential of yours changes hands, because nothing in this path ever needs one.
You ask, in plain language
Which tables hold the subscriptions, what is in this one, write me the query. The assistant goes and looks, mid conversation, rather than asking you to paste the schema in again.
The standard
It runs on MCP, which your assistant already speaks.
MCP, short for Model Context Protocol, is the open standard an assistant uses to reach something outside itself and read from it. Claude and Claude Code speak it natively, and the list of clients that do keeps growing.
What that means in practice is that nobody writes an integration. You add Taavik as an MCP server once, in whatever client you use, and from then on the assistant decides by itself when it needs to look something up. There is no plugin to maintain and no API for you to call.
Connecting it
One URL, pasted once.
Every client handles connectors slightly differently, but they all want the same thing from you: the address of the server. This is it.
https://api.taavik.com/mcp
Before you start
You need a Taavik account and at least one database connected, because the connector reads the catalog the agent produced. If your first scan has not finished yet, there is nothing for the assistant to look at.
In your client
Open the connector settings, add a custom server, and paste the URL above. A browser window opens so you can sign in to Taavik and approve the access. After that the assistant handles the rest on its own.
You sign in to Taavik, not to your database. The approval covers reading the catalog and nothing else, and you can revoke it from the workspace whenever you want.
MCP setup, per client
Claude Code, Cursor, Windsurf: same URL, three configuration files.
Every client that speaks MCP takes the same server URL. What differs is the file the client reads it from, and the name it assigns to the connector inside its own UI. The three examples below cover the clients asked about most often; other MCP clients follow the same shape with different file paths.
Claude Code
Add Taavik as a custom MCP server through Claude Code settings, or drop the connector block into .claude/mcp.json at the workspace root:
{
"mcpServers": {
"taavik": {
"url": "https://api.taavik.com/mcp"
}
}
}Cursor
Open the Cursor settings, find the MCP servers panel, and add a new HTTP server with the URL above. Cursor also reads a workspace-level .cursor/mcp.json file with the same block:
{
"mcpServers": {
"taavik": {
"url": "https://api.taavik.com/mcp"
}
}
}Windsurf
Windsurf uses the same MCP shape through its Cascade panel. Add a custom server pointing at the URL, confirm the sign-in in the browser window that opens, and the connector becomes available to Cascade and to Windsurf chat.
{
"mcpServers": {
"taavik": {
"url": "https://api.taavik.com/mcp"
}
}
}The sign-in happens once, in the browser window the client opens on the first connection. From then on, the assistant reaches the catalog on its own whenever it decides it needs it.
One connector, every engine
MCP for PostgreSQL, MySQL, SQL Server, and Redshift.
The connector reads whatever the agent scanned, and the agent knows how to scan all four engines. A workspace with three MySQL connections and one PostgreSQL connection exposes them through the same MCP endpoint, and the assistant lists them, picks the one the question is about, and reads its tables from there. There is nothing per-engine to configure on the client side.
PostgreSQL
MCP database context read from the agent's PostgreSQL scan. Tables, columns, foreign keys, indexes, views and routines are all visible to the assistant through the same tool set.
MySQL
MCP for MySQL through the agent's information_schema scan. Provider quirks (utf8mb4, backtick identifiers, generated columns) are normalized before the catalog is exposed to the assistant.
SQL Server
MCP for SQL Server via the agent's sys.tables and sys.columns scan. The assistant sees schemas, table types, computed columns, and check constraints the way SQL Server exposes them.
Amazon Redshift
MCP for Redshift through the agent's catalog scan. Distribution styles and sort keys travel with the table definition, so the assistant can reason about the shape a query will hit.
The difference
Two things a direct database connection cannot give it.
Both are knowledge your database does not contain, which means anything that plugs straight into it is missing them by construction rather than by oversight. That is not a gap a competitor can close by shipping faster.
The relations nobody declared
Your team knows this column points at that table. The database was never told. Draw the relation on the schema map once and the assistant reads it from then on, alongside the real foreign keys, which is human knowledge finally written down somewhere a machine can reach.
The values a column actually holds
When a column carries a short list of values, a status, a type, a country code, Taavik records which ones are in there and how often each appears. So the assistant filters on COMPLETED because that is what your data says, not on the spelling it would have picked on its own.
The boundary
What the assistant reaches, and what it never does.
Everything the connector returns comes from the catalog Taavik already holds, so asking a question never opens a connection towards your database. There is nothing there for the assistant to reach, even if it went looking.
It can
- List the databases connected to your workspace
- List their tables and views
- Read one table in full: columns, types, nullability, defaults, keys, indexes
- See the relations you drew and the tables pointing back at this one
- See the value ranges of short columns, while that setting is on
It cannot, at all
- Receive your database credentials, at any point
- Open a connection towards your database
- Run a query against your data on its own
- Write, alter or delete anything. Every tool is read-only
- Act unrecorded. Each call is stored with the user who made it
The full picture of what leaves your network, and in which two cases your tables are read at all, is on the security page.
Pricing
Schema access is on the free plan.
Connect a database, add the connector, and your assistant knows your schema without anyone reaching for a card. What starts at Pro is the part that stops it guessing values, along with workspace details and audit reading, and that first one is usually why teams move up.
FAQ
MCP database connector: give your AI assistant your real schema
Does my assistant get the database credentials? +
No. The connector reads the catalog Taavik built from your scans. The credentials stay sealed for the agent inside your network, and nothing in this flow needs them.
When I ask a question, does a query hit my database? +
No. The answer comes from the catalog already stored in the cloud, so it arrives immediately and puts no load on your database. Your database is read by the agent on a scan, and when you yourself run a saved query.
Which assistants can use it? +
Any client that speaks MCP, the Model Context Protocol. Claude and Claude Code support it natively, and the list of clients keeps growing. Taavik is the MCP server, your assistant is the client.
Is this a normal MCP server? +
Yes, with one difference worth knowing. Most database MCP servers connect straight to your database using credentials you hand over. This one reads the catalog Taavik already scanned, so there are no credentials to hand over and no route from the assistant to your data.
Can the assistant change something by accident? +
No. Every tool is declared read-only and none of them writes. The worst case is an answer you disagree with, not an altered database.
How up to date is what the assistant sees? +
As up to date as your last scan, and no more than that. A table renamed this morning shows up once the next scan has run, which is the honest limit of the approach, but there is no export to redo and no file to upload again.
What is included on the free plan? +
Listing your connected databases, listing tables and views, and reading a full table description. Value ranges of columns, workspace details, member lists and audit reading are on Pro.
Is every request recorded? +
Yes. Each tool call is stored with the user who made it and what was asked, in the same audit trail as scans and query runs.
Stop pasting your schema into a chat window.
Free for one connection. Install the agent, let the first scan finish, then point your assistant at the workspace and it will look things up on its own from that moment on.