Skip to content
For agents

Ask your website, not a dashboard

Conseto speaks MCP, so the assistant you already use can read this site's traffic, consent and website check data directly. No export, no copy and paste, no second tool to learn.

What it is

MCP (Model Context Protocol) is how an AI assistant plugs into a data source. You paste one address and one key into Claude or Cursor, and from then on you ask questions in your own words. Conseto answers from the same numbers the dashboard shows, because it runs the same query underneath.

  • Read only. An agent can look at your data, never change it, never delete it.
  • One key, one site. A key opens exactly one project and nothing else.
  • Bot traffic is already excluded, and every answer says how much was removed.

How to connect

  1. Open Settings, Projects, the site you want, and find Agent access (MCP).
  2. Create a key. It is shown once, so copy it before you close the page.
  3. Paste the block for your tool.

Claude Code

Run in a terminal.

bash
claude mcp add --transport http conseto https://api.conseto.io/mcp \
  --header "Authorization: Bearer cmk_YOUR_KEY"

claude.ai

Add a connector for https://api.conseto.io/mcp, sign in, and pick the project.

Cursor

Into .cursor/mcp.json in your project, or ~/.cursor/mcp.json for every project.

JSON
{
  "mcpServers": {
    "conseto": {
      "url": "https://api.conseto.io/mcp",
      "headers": {
        "Authorization": "Bearer cmk_YOUR_KEY"
      }
    }
  }
}

Better: keep the key out of the file. Put "Authorization": "Bearer ${env:CONSETO_MCP_KEY}" in its place and export CONSETO_MCP_KEY in your shell. A key in .cursor/mcp.json is a long-lived secret in a file that gets committed.

Claude Desktop

claude.ai works today over OAuth sign in. Claude Desktop uses the same kind of connector and should work the same way, but we have not tested it yet.

What the agent can do

Eight tools on the Business plan and up. The agent picks between them on its own, you never name them.

describe_model
Lists everything this project can be asked about: sources, metrics with their units, dimensions with their value formats, and how far back your plan keeps data. The agent calls this first so it can never invent a metric.
query_metrics
The numbers: any metric, broken down by up to two dimensions, over any period, as a time series or against the previous period. Returns the rows and the exact SQL behind them.
list_records
Individual records instead of totals: one visitor's path, one visit, one order. For when the aggregate is not the answer.
top_values
The most common values of a dimension, with counts. This is how the agent learns that your channels are called Direct and Organic Search before it filters on them.
list_projects
Which project this key opens, its plan and its retention window.
list_decisions
The project's decisions: what needs your call and what was already decided, each with the evidence it stands on (a query you can run with query_metrics), its state and, once it is done, the measured result before and after. By default the active ones. Read only: you accept, complete or dismiss a decision in the Conseto app.
get_decision
One decision by its id (from list_decisions), with a task brief in markdown: the project, what is going on, the numbers, the evidence query, the goal and how success is checked (Conseto compares the evidence metric over the 7 days before and after the decision was marked done). Read only.
ask
Ask this project's data a question in plain language and get a short answer, plus the query traces it is grounded in (the same numbers the dashboard shows). Takes a question up to 500 characters and an optional project ID, checked only as a safety net against the key's own project, never a route to another one. Business plan and higher.

Questions that work

Written the way you would actually say them.

  • How many people came from ChatGPT in the last 30 days?

  • What is our consent rate, split by legal regime?

  • Which Google queries rank in the top 3 but get no clicks?

  • Compare mobile and desktop visits to the same month last year.

  • Show me every page that lost traffic this week.

  • What did the last security scan find?

Limits

  • Business and Enterprise plans. Free and Starter do not include MCP.
  • 30 requests per minute per key.
  • Questions reach as far back as your plan retains data. Older ranges are trimmed to what exists, not refused.
  • Up to 10 active keys per project, so each machine can have its own.
  • ask has a separate budget: 200 questions a month on Business, 1,000 on Enterprise, shared by every key on the project and reset on the 1st (UTC). Every answer says how many remain.

Security

  • Read only, at the protocol level. There is no tool that writes, and no way to ask for one.
  • A key is bound to one project. It cannot see your other sites and cannot see anyone else's.
  • We store only a fingerprint of the key, never the key. That is why it is shown once.
  • Revoke a key in settings and it stops working immediately, without touching the others.
  • Raw event parameters are not available over MCP. They are free-form data written by your own site, so we cannot promise what is inside them.

Your data, in the tool you already have open

MCP is included in Business. Create a key and connect.

Read next

Checked against the code on 1 October 2026.

Missing something?

This is everything we document today, and each page says only what the script and the gateway do now. If you need a guide that is not here, tell us what you want to achieve.

Write to us