Feature

AI agents in the canvas

Brief the agent like a colleague. It builds tables, applies filters, derives variables, and writes analysis — directly in your workspace, on your data, with the model you choose.

Surface parity, not chatbot bolt-on

Every screen in Recense implements a single Surface protocol: a canonical Document of intent, an Action vocabulary for changes, and a Snapshot of observable state. Agents read and write through the same protocol the UI does. There is one notion of truth in the canvas; both human and agent operate it.

Practically: you can ask an agent to build a brand-tracking crosstab by region with sig testing at 95%, and the result is a real table in your canvas — editable, filterable, derivable — not a screenshot pasted into a chat thread.

The Recense canvas with an agent thread alongside a crosstab the agent built.
One canvas, one notion of truth. Human and agent operate the same surface.

What an agent can do

  • Build crosstabs and pivots, set measures and weights, apply filters.
  • Derive variables (recodes, nets, indices) and add formula cells.
  • Search variables, summarise the dataset, find related questions.
  • Draft narrative analysis next to the tables it built.
  • Iterate on a brief — add cuts, refine cells, challenge weak findings — without losing context.
  • Hand off between specialists: a planning agent, an analysis agent, a storytelling agent.
An agent producing tables, derived variables, and narrative analysis in a single thread.
Crosstabs, derives, search, narrative — all visible in the canvas.

Bring your own model

Frontier capability changes month to month. Locking a tool to one embedded model is a tax on the work. Recense supports two paths:

  • Bring your own key (BYOK). Connect your Anthropic, OpenAI, Google, or Fireworks key. Conversations go from your browser to the provider — Recense never sees them.
  • Built-in agent. Use a managed agent with included credits if you'd rather not handle keys. Provider choice is exposed; we proxy and don't store conversation text.
  • MCP integration. The Recense canvas surfaces are exposed over the Model Context Protocol. Use Recense from Claude Desktop, Cursor, or any MCP-aware client.
Provider picker showing Anthropic, OpenAI, Google, and Fireworks options with key configuration.
Frontier capability changes month to month. Don't be locked to one model.

Transparency by default

Every agent action is logged as a tool call against the canvas Surface. You can see what the agent built, why, and re-run or revert. There is no hidden sub-agent computation, no black-box "AI Insights" panel — just visible, undoable changes you can inspect.

A timeline of agent tool calls against the canvas surface, each with revert and re-run controls.
Every agent action is a visible, undoable change.

Brief an agent on your data

Sign up, drop in a .sav, and put an agent to work.

Get started