Skip to content

Using u-Slicer MCP

You talk to the assistant, not to the tools: the model chooses which tool to call and with which arguments. This page describes what it has to work with, so you can steer it and understand its answers.

Available tools

list_projects
Lists the projects you have access to.
list_slicers
Lists the slicers of a given project.
slicer_info
Returns a slicer schema: its dimensions, measures, available date range and defaults, including the maintainers' notes on what a column means.
query
Runs a table query: dimensions, measures, filters, ordering and a period. Equivalent of the table in the UI.
time_series
Returns the evolution of one measure over time, optionally split by dimension values and compared against the total. Equivalent of the chart in the UI.
uslicer_url
Builds a u-Slicer UI link for a described view, without running the query.
query_url
Takes a u-Slicer UI link (full or short /<project>/go/<code>), decodes it, runs the underlying query and returns the data.

The tool list and their arguments may change as the server evolves. Assistants discover them automatically, so this normally needs no attention.

Note

The tools mirror the APIv2 methods — query, time_series, info, url_to_api. Filters, date formats and ordering follow the same rules as the API.

What a query supports

Both query and time_series accept:

  • Period — start_date and end_date in YYYY-MM-DD format, plus an optional timezone offset in hours.
  • Dimensions — the categorical columns to split by.
  • Measures — the numeric columns to compute. time_series takes exactly one.
  • Dimension filters — conditions on dimension values, combined with AND.
  • Measure filters — conditions on measure values after the final aggregation.
  • Currency conversion — a measure whose semantic type is money can be returned in another currency (ISO 4217 code).
  • Ordering and a row limit (query), and the dimension used as the X axis (time_series).

So in practice you can ask for anything you would build in the UI, in one sentence:

Using u-Slicer, in project <project> and slicer <slicer>, show revenue and impressions by country for last month, only for countries with more than 1M impressions, ordered by revenue, top 20, in EUR.

Reading the answer

query and time_series return the data rows together with:

  • url — the proof link, opening the same data in the u-Slicer UI. Use it to confirm the assistant queried what you meant.
  • confidence_ranges — present only when the data is sampled or compressed. A confidence range of 0.03 on a row means its measure values may deviate by about 0.03% from exact values; smaller is more precise. When the field is absent, the data is exact.

Dimension values may be shown with an extended display name in parentheses. The display name is not part of the value itself.

If you already have a u-Slicer view, paste its link into the chat:

What does this report show, and how does it compare to the previous month? https://uslicer.criteo.com/<project>/go/<code>

The assistant uses query_url: the link is decoded, relative periods such as "last 30 days" are resolved to absolute dates, and the query is executed. It can then re-run variations of that query for you.

Getting better answers

  • Name u-Slicer in the prompt. It tells the model to reach for the tools instead of answering from memory.
  • Name the project and the slicer when you know them. Otherwise the assistant spends turns listing and guessing.
  • State the period explicitly. "Last week" is ambiguous; 2026-08-17 to 2026-08-23 is not.
  • Ask for the schema first when exploring an unfamiliar slicer: "which measures exist and what do they mean?" The notes stored on the columns usually answer it.
  • Keep requests narrow. Query performance depends directly on the date range, and a request with many columns fills the model's context with data instead of analysis. Ask for the top N rows rather than everything.
  • Check the proof link before acting on a number.

Limits and expectations

  • Read-only. The tools query data. They do not create columns, dashboards, bookmarks or scheduled exports.
  • Your permissions apply. The server does not elevate privileges: it validates your token and forwards it to the u-Slicer API.
  • The model can be wrong. A wrong column, a shifted period or an over-confident summary are all possible. The proof link is the check.
  • Large results are not the goal. Very wide or very long tables are expensive to query and hard for a model to reason over — narrow the question instead.