DatoricDatoric

Solutions / Computer-use agent

Computer-use agent
data

Human demonstrations that align screen states, UI actions, structured context, and reviewed outcomes. Built to the schema and environments your agent needs.

ASCII illustration of an open laptop computer
Available inventory
250,000 traces
Screen activity
Approximately 15,000 hours
Interaction capture
Screen states + UI actions
Structured context
DOM + accessibility

01

Agents need aligned trajectories, not screen recordings

Computer-use agents learn from demonstrations that align screen states, actions, structured interface context, and reviewed outcomes. Screen video alone cannot provide that grounding.

Explore the details

Deployable data also needs task and environment coverage, reviewable rights, and QA records that match the applications and workflows an agent will encounter.

ASCII illustration of arrow and hand mouse cursors

02

What each trajectory contains

A usable trajectory connects what the screen showed, what the user did, and the interface state behind it with an explicit goal and reviewed outcome.

Aligned signal types
  • Screen recordings: Full-resolution capture of the screen while a task is performed.
  • Screenshots and screen states: Per-step still frames captured before and after each action.
  • UI action logs: Structured, timestamped events for clicks, typing, scroll, and keys.
  • DOM and accessibility-tree snapshots: The structured page state behind each screenshot.
  • Task and goal metadata: The instruction, category, and environment for each trajectory.
  • Success labels and QA review: Outcome labels and review fields aligned to each trajectory.

03

Coverage follows where your agent runs

Start with the available computer-use dataset, then scope environments, task shapes, languages, fields, and evaluation criteria around the deployment.

Coverage dimensions
  • Browser applications, desktop operating systems, and native apps.
  • Productivity suites, email, CRM, data entry, and admin consoles.
  • Single-step actions, multi-screen workflows, forms, search, and navigation.
  • Dynamic content, permission states, recoverable errors, and retries.
  • Multiple locales and languages, including cross-application task chains.

04

Grounded trajectories you can audit

Each record anchors a goal to ordered steps with typed actions, resolvable targets, aligned observations, and a reviewed outcome.

Explore the details

QA can include completion review, PII redaction, duplicate filtering, malformed-log rejection, contributor consent records, chain of custody, and commercial license scope.

ASCII illustration of a desktop interface with a large mouse cursor

05

From sample to scale

Validate the schema and review evidence on a representative sample before committing to a larger collection.

Program workflow
  • Scope: We align on task categories, environments, locales, modalities, and the trajectory schema your training or evaluation pipeline expects.
  • Sample: You receive a representative review package to validate against your loader and the evaluation checklist before any volume commitment.
  • Pilot: A bounded batch is collected to the agreed specification, with labels, QA, and redaction requirements applied so you can evaluate it on your own tasks.
  • Scale: Once the pilot clears your acceptance bar, collection scales to the target volume and cadence with the agreed QA and provenance records.

06

Choose data that fits the deployment

Evaluate schema fit, environment coverage, labels, QA, provenance, and commercial rights together. A headline trace count cannot answer those questions.

Questions worth asking any provider
  • Do actions, observations, timestamps, and targets align step by step?
  • Which browser, desktop, application, locale, and workflow categories are covered?
  • Are completion outcomes human reviewed, and which malformed records are rejected?
  • How are duplicates, PII, contributor consent, and chain of custody handled?
  • Can your team inspect a real sample and schema before approving a pilot?

Program standards

Review how the data is sourced, checked, and licensed.

Data rightsQualityLicensing

Related datasets

Agentic

Computer-Use Traces Dataset

Available · 250,000 traces

Keep reading

Guide

Computer-use agent datasets

A buyer guide to trajectory structure, coverage, labels, QA, and rights.

Guide

Commercial vs open GUI-agent data

Where public releases help, and where production programs need more control.

Tool

Computer-use schema inspector

Check a trajectory sample against the published reference fields in your browser.

Research

VidWork-Bench

Datoric's evaluation of frontier video models on workplace task understanding, the closest research signal to how agents read on-screen work.

Frequently asked

Questions buyers ask

What is a computer-use agent dataset?

It is a collection of task trajectories showing how an interface is operated to reach a goal. Records can pair screen states with timestamped UI actions, structured context, instructions, and reviewed outcomes.

Why are DOM and accessibility-tree snapshots important?

They expose element roles, names, values, and hierarchy. That structure supports semantic targeting alongside pixel coordinates and helps teams evaluate behavior across interface variations.

Can Datoric collect data for desktop applications, not just the browser?

Yes. The available dataset covers desktop applications, browsers, and productivity and enterprise software. Custom collection can target additional environments, tasks, fields, and evaluation criteria.

How is quality evaluated?

The specification includes human-reviewed completion labels, PII redaction review, duplicate filtering, and malformed-log rejection. Buyers can review samples, metadata, QA summaries, and licensing records before scaling.

How do we start?

Request a sample. We scope tasks, environments, modalities, and schema, deliver a representative package to validate, run a bounded pilot with labels and QA, then scale to the target volume.

Start with a computer-use sample

Tell us the tasks, environments, and schema you need. We will follow up with a representative sample, then scope a pilot.

Browse datasets