Start here

What arkush is and is not

Arkush makes dashboards on a free canvas, where every chart reads saved rows. This page says what arkush does, what it does not do, and why, so you know before you build on it.

What arkush is

• A free canvas. You put each chart, table and note where you want it, at the size you want. No grid of widgets decides the layout.
• One document per dashboard. Everything a dashboard holds is one JSON document: the blocks, the queries, the saved rows, the theme. You, the Source panel and your AI agent edit the same document.
• SQL in, Vega-Lite out. Data arrives through SQL, and every chart is a Vega-Lite chart: Vega-Lite is an open format that describes a chart as JSON. Both are open languages, so what you learn here works elsewhere.
• Saved rows for every reader. Each datasource keeps the rows of its last run. Readers read those rows, so a reader never runs a query and never costs anyone a warehouse bill.
• Made for your own AI agent. An agent such as Claude Code connects over MCP, the open protocol that AI assistants use to reach tools, reads the dashboard, and edits it under your account. See Connect your AI agent.

Data and the warehouse

Opening a dashboard runs no query

A dashboard shows the rows from the last refresh of each datasource. There is no live connection to BigQuery while someone reads. Because of this, ten readers cost what one reader costs, readers need no warehouse access, and a page opens the same whatever the warehouse is doing. To get new numbers, open the Data panel and select Refresh on a datasource. Each datasource shows the time of its last refresh.

On arkush.app, BigQuery runs under your own Google account

A BigQuery query runs in your browser, under your Google account, when you select Refresh. It bills a Google Cloud project that you choose, so you need a Google Cloud project of your own. If you signed in with GitHub, you connect a Google account for BigQuery as a separate step: see Query BigQuery with your own account.

The arkush.app server holds no warehouse access of its own. It never runs your query, and no schedule refreshes it on arkush.app. An AI agent works from the rows of your last refresh. A query that refreshes on a schedule exists only on an org install, the arkush that an organization runs for its own people: see What differs on an org install.

Without an account, nothing connects to Google

Without an account, you can import files, read a Google Sheet that is shared by link, and build derived views. Your browser does all of this work, and the server does nothing for a person without an account. To connect BigQuery or a private sheet, sign in.

A datasource holds at most 50,000 rows

Charts draw in the browser, and the limit keeps every dashboard fast to open. A billion-row table is never read whole. Aggregate it in the warehouse to the level of detail that the dashboard shows: group and filter in SQL, and keep every column that a control filters on. A BigQuery result over the limit keeps only the first rows, and a larger file opens a step that reduces it with SQL. See Bring data in.

Filters slice the saved rows

A control or a click filter narrows the saved rows in the reader's browser. It never sends a new query, so a reader cannot ask for rows that the datasource does not hold. To change what a datasource fetches, change its SQL or its query parameter, and refresh it. Fetch broadly enough for the filters, then let the filters slice.

Every reader reads the same rows

A dashboard cannot show each reader a different set of rows. A filter per reader would need a live query for each reader. To show different rows to different audiences, make one dashboard per audience, each with SQL that already filters to that audience.

BigQuery is the one warehouse

Arkush queries BigQuery. On arkush.app, no datasource reads a database, so data from another warehouse or a database reaches arkush.app as a file: export it to CSV or Parquet, and import it. More data sources are one of the directions.

An org install can also read a PostgreSQL database on Cloud SQL that the people who run it register: see Query a database. There, the query runs as a service account. Arkush never stores a password that opens a database, so a database that takes only a password reaches arkush as a file, or through BigQuery.

Your browser reads addresses from Google's hosts only

A Google Sheet that is shared by link, or a file stored with Google, reads straight from its address into your browser. A CSV on another host does not. The app lets the page talk only to the hosts it needs, so that a fault in the page cannot send your data to a site it does not know, and each added host would widen that list. Download the file and import it.

SQL and Vega-Lite are the languages

Arkush has no formula language of its own. Write the logic in SQL, or in a derived view over the other datasources. A measure holds one named calculation, such as a ratio of two sums.

No semantic layer

Arkush has no metrics store and no modeling step. The SQL is the model. Labels, formats and measures on a datasource carry the meaning of each column.

A derived view is saved, not live

A derived view keeps its result like any other datasource. A chart filter cannot reach a column that the view grouped away. Group the view by the chart's level of detail plus every column that a filter needs.

A server run stops at 30 seconds and 1 GB

Some datasources run on the server: a derived view that an AI agent runs, and a datasource that reduces a file kept in the files store (see Store a file for reuse). Such a run stops after 30 seconds or 1 GB of memory, so one run cannot hold the server for everyone. Filter before you join, or select fewer columns.

Sharing and reading

On arkush.app, a dashboard is private or published

On arkush.app, you cannot share a dashboard with a named person or open it to everyone signed in, and there is no way to ask for a share. A share would show strangers each other's email addresses. Only the people who run arkush.app add shares. Publish the dashboard to the internet instead. See Share a dashboard.

A published dashboard is a frozen copy

The published page changes only when you publish again, so readers never see it change under them. It shows no comments and no email address: a generated username names the author. Only the owner publishes. To show your edits, select Update published copy.

One chart is not published alone

Each chart of a published dashboard has its own page, but you publish the whole dashboard. To publish one chart, copy it into a dashboard of its own and publish that dashboard.

Anyone who can read a dashboard can take its rows

The reader's browser holds the rows to draw and filter the charts, so a reader can save them. When the detail is sensitive, aggregate it in SQL before it reaches the dashboard.

The canvas does not reflow for a phone

A free canvas has no correct reflow, and an automatic one gives layouts nobody chose. A phone opens a dashboard in view mode, scaled to the dashboard's design width, and you pinch to zoom. The author sets the design width in Dashboard settings.

No embedding

Embedding a dashboard in another page is not available. Link to the published page.

Building

Without an account, dashboards live in this browser

Nothing backs them up, and clearing the browser deletes them. Sign in to save them to your account.

No plugins, raw HTML or embedded pages

A plugin block would break the shared look, the JSON that an agent reads, and the checks. Raw HTML needs a filter that can fail. Use the built-in blocks and Vega-Lite.

An import is a copy, never a backup

Import keeps the data, the datasources and the layout. It leaves out the sharing, the comments, the scheduled sends of an org install, and the stamp that says how far to trust the dashboard, because the file can come from anyone. Keep your own export of what you cannot lose: open the dashboard, open the Dashboards page, and select Export open dashboard… at the foot of the page.

No drill-down or saved views

A click or a drag on a chart can filter the dashboard, but it does not open a level of detail. A reader cannot save an adjusted view. Use controls and chart filters, and send a link: the link carries the filters.

A shared datasource is a copy

The datasource library is a set of ready-made datasources that people with the analyst right keep for everyone on a server. A datasource from the library is copied into your dashboard and does not follow later changes to the library. A live link would let one library edit break every dashboard that reads it. To get a change, add the datasource again.

On arkush.app, the analyst right belongs to the people who run it, so the library holds only what they put there.

AI agents

No AI chat inside the app

Arkush is what your agent edits. It does not bundle an assistant of its own. Connect the agent you already use, such as Claude Code, over MCP.

An agent cannot run a query under your Google account

An agent never holds your Google sign-in. It works from the rows that you last refreshed. Refresh in the app, then ask the agent.

An agent does not widen access

An agent cannot publish a dashboard or change who can open it. It runs no warehouse SQL of its own: every query it writes is saved in a dashboard's datasource, where you see it before anything runs. Publish yourself. An agent explores the saved rows instead.

An agent follows only its own person's comments

An agent carries out a comment that you wrote. On a dashboard that other people can comment on, a comment from anyone else is a request that the agent weighs, so another person cannot steer your agent. On arkush.app, other people comment only on a dashboard that the people who run it shared.

Working together

This section applies to a dashboard that several people edit. On arkush.app, that is a dashboard that the people who run it shared, or one dashboard open in two of your own tabs.

No live co-editing of the same text

People who edit one dashboard see each other on it, and changes to different parts merge. When two edits collide, the other edit wins. The app sets your version aside in the dashboard's version history, and the set-aside version stays for 30 days. Open Version history in Dashboard settings to take it back.

No email sends and no alerts

Arkush does not email a dashboard and does not alert on a threshold. On arkush.app, it does not send a dashboard anywhere. An org install sends a dashboard to Slack on a schedule: see What differs on an org install. Email and alerts are among the directions.

Privacy

The software reports nothing home

The arkush software sends no telemetry, no error reports and no usage statistics to anyone. The people who run a server keep its logs: the privacy policy of arkush.app says what its logs hold.

Agents never hold your Google credential

An agent gets its own access tokens, which work on arkush only. Every change an agent makes carries your name.

The limits in numbers

For where arkush can grow, read Directions.