Vecten MCP

Ask your company's data from Claude, as yourself

Your people ask questions from the AI assistant they already use. Your warehouse decides what each person can see, and every answer comes with a receipt. It works with the warehouse you already have and takes hours to set up.

How it works

One question, start to finish. Pick who's asking.

MCP is the open standard Claude and other assistants use to reach data. Vecten MCP sits between the assistant and your warehouse, inside your own cloud account. No question or answer passes through Vecten.

Who's asking?

What was revenue last quarter, by region?

Revenue last quarter, by region
Europe€4.2M

Other regions aren't in your access.

Receipt
Asked by
Anna, sales lead for Europe
Definition
Revenue: orders minus refunds
Data
Orders and refunds, last quarter
Access
Read-only, as Anna
  1. You

    Ask in your AI assistant

    Signed in as Anna, sales lead for Europe.

  2. Vecten MCP

    Use the agreed definition

    Revenue means orders minus refunds, the same for everyone.

  3. Your warehouse

    Run it as that person, read-only

    Anna can read Europe only, so that's all the warehouse returns.

  4. Vecten MCP

    Answer with a receipt

    The receipt shows who asked, which definition was used and which data.

  5. Vecten MCP

    Record it for admins

    Admins can see who asked what.

Same question, three people, three answers. The AI never decides who sees what: your warehouse does, using the access rules you already manage.

01

Ask

A person asks a question in Claude Desktop, Claude Code or any other client that speaks MCP. The first time, they sign in with a password or through your identity provider.

02

Run as that person

The server runs the query under that person's own warehouse identity, read-only, with a row cap and a timeout. What their permissions forbid, they cannot read.

03

Keep the receipt

The answer comes back with a receipt. The call is recorded with the person, the question, the SQL and the tables it touched, and anyone signed in can check the receipt later.

Secure by design

The warehouse decides who sees what. The AI never does.

01

Everyone asks as themselves

One identity per person

Each person gets their own warehouse account, and every question runs under it. The access rules you already manage in the warehouse apply to the AI too.

Identity and access (opens in a new tab)
02

Read-only, with limits

Nothing gets changed

Questions can only read data, never change it, and each one has a size and time limit. Anything an AI writes is treated as untrusted input.

MCP tools (opens in a new tab)
03

No master password on the server

Your administrator reads the setup first

The server holds no admin access to your warehouse and keeps nobody's password. The setup steps are printed for your warehouse administrator to read before anything runs.

What it does not protect against (opens in a new tab)
04

One definition of each number

Agreed once, reviewed like code

An AI left to itself will invent its own version of revenue. Here each number is defined once and reviewed, every answer names the definition it used, and when the data can't answer exactly, it says so instead of guessing.

Authoring the semantic model (opens in a new tab)
05

A record of every question

Who asked what

Admins see who asked what, which questions couldn't be answered and which definitions are missing. The record lives in the server's own private database, never in your warehouse.

Audit records (opens in a new tab)
The learning loop

A wrong answer gets fixed once. For everyone.

When someone spots a wrong number, the report doesn't disappear into a ticket queue. It becomes a tracked gap, and the fix comes back as a change a person reviews, with a test so the mistake can't quietly return.

From a wrong answer to the right one

  1. Anyone

    Flags a wrong answer

    “Revenue should exclude refunds.” They report it right from their AI assistant.

  2. Vecten MCP

    Tracks it as a gap

    Similar reports and failed questions are grouped, so admins see what's missing.

  3. AI agent

    Proposes the fix

    A corrected definition, plus a test built from the original question.

  4. A person

    Reviews it

    Nothing changes until someone approves it.

  5. The test

    Guards the fix

    If a later edit brings the mistake back, the check fails.

  6. Everyone

    Gets the right answer

    The next person who asks gets the corrected number.

Every report makes the next answer better, back to “Flags a wrong answer”.

Under the hood

Use it on the warehouse you already have. Leave Core out.

If you run Postgres or Snowflake and your assistants cannot reach that data yet, the MCP serves it without Core. The commands below set up a project against an existing warehouse. The last one lists what the project needs, in order: the setup SQL for your administrator, the semantic model, the first people.

shell
uv tool install vecten
vecten init warehouse-assistant --name warehouse-assistant --users 5 \
  --has-warehouse yes --warehouse-url postgres://warehouse.example.com/analytics
cd warehouse-assistant
vecten next

The warehouse address never holds a password. Once the server is deployed, people add its address as a custom connector in Claude Desktop, or with one command in Claude Code. The MCP's own README (opens in a new tab) walks through it.

Warehouses
Postgres and Snowflake. Databricks, BigQuery and Redshift are planned.
Runs on
Your laptop with Docker (Postgres only), or your account with major cloud providers like AWS.
Snowflake
Deployed only. Snowflake trusts the server's sign-in at its public address, so a laptop cannot stand in.
Clients
Claude Desktop, Claude Code and any client that speaks MCP over streamable HTTP.
Metrics
One dataset per metric query. A question that needs a join gets a SQL starting point instead.
THE NEXT LEVEL

Run it yourself, or let us run it for you.

Vecten Open is free and complete. It runs in your own cloud account, nothing passes through Vecten, and you don't need a partner or a paid plan. If you stop working with us, you keep everything: the code, the data, the definitions and the record of every question. When you'd rather not run it yourself, or need it built, connected and compliant, that's the work Vecten does.