We record real work, the way a model needs to see it.
Egocentric cameras on tradespeople and nurses, screen recorders on bookkeepers and support agents, trained operators on robot arms, and specialists working through tasks in their own field. Every recording is consented, reviewed and delivered licensed.
Six ways to record the real thing.
Each method is run by the people who do that work every day, to a written protocol, and reviewed before it reaches you.
01Egocentric video
Tradespeople, cooks, nurses, mechanics, warehouse and farm workers wear a small head- or chest-mounted camera through their normal shift. The footage shows a task from the worker's point of view: where they look, what they pick up, how they test and when they decide it's done.
Yields: First-person video with task and phase labels, hand and tool annotations, consent on file.
02Screen & workflow recording
Bookkeepers, support agents, designers, editors and office workers record only the apps they choose while doing their real job. No camera. The recordings teach computer-use agents how work actually gets done across the tools people use every day.
Yields: Screen video with click, keystroke and window events, task boundaries, personal data masked.
03Robot teleoperation
Trained operators drive robot arms and mobile manipulators to a written task protocol: initial conditions, object sets, allowed strategies and success criteria, so episodes are comparable across operators and sessions.
Yields: Joint states and actions, end-effector poses, wrist and scene video, success and failure labels.
04Voice & conversation
Speakers, call handlers and linguists record scripted and natural speech, service calls and multilingual dialogue in real acoustic conditions, with every participant's consent recorded before the first take.
Yields: Audio with verbatim transcripts, speaker and language tags, accent and environment metadata.
05Expert demonstrations
Physicians, lawyers, engineers, scientists and other specialists work through tasks from the open opportunities: writing and grading answers, reasoning traces, comparisons and reviews in their own field, at the rate shown before they accept.
Yields: Expert-written demonstrations, preference pairs, rubric scores and evaluation sets, second-expert reviewed.
06Sensor & environment capture
Where a model needs more than video, we add what the task calls for: depth and 3D scans of rooms and objects, force and tactile signals on robot hardware, vehicle and equipment telemetry, synchronized with the camera streams.
Yields: Depth maps, point clouds and meshes, force and tactile streams, telemetry aligned to video.
One protocol, start to finish.
The same four steps run every capture project, whether it's ten electricians wearing cameras or a thousand hours of screen work.
Brief & protocol
You tell us the capability. We write a capture protocol: who records, which tasks, what the camera or screen must show, variations to cover and what counts as a complete episode.
OutputWritten capture protocolTurnaroundUsually within daysRecruit & consent
We recruit the people from the capture programs and the expert network, verify their identity and credentials, and collect consent from everyone who will be recorded, including employers and customers on site.
WhoVerified workers and expertsConsentSigned before the first recordingRecord
People record their real work with the kit we send, or on the apps they already use. Our team checks the first sessions against the protocol before anyone records at volume.
Pay$12 to $30 per accepted hourEquipmentSupplied by us, or your hardwareReview & deliver
Every recording is reviewed for protocol adherence, faces and screens outside scope are masked, each file is fingerprinted, and the accepted set is delivered with labels, a signed provenance record and a receipt by hash.
QualityReviewed episode by episodeDeliveryWebDataset, Parquet, JSON Lines
Teaching agents how work gets done on a computer.
Office workers record the apps they choose, and nothing else, while doing their normal job. Clicks, keystrokes and window changes are captured alongside the video, with personal data masked before review.
Consent before capture
Nothing is recorded until the worker, their employer and anyone on site has agreed in writing. Each release is bound to the episodes it covers.
Real work, not staging
People record the job they already do, during their normal shift. Variation comes from the protocol and from real conditions, not from a set.
Masked by default
Faces, screens and documents outside the task are blurred, personal data in screen recordings is masked, and location metadata never leaves intake.
Traceable to the frame
Every file is fingerprinted and frame-hashed at intake, so duplicates are caught and a model checkpoint can be traced to the exact recordings behind it.
Before you ask.
Who does the recording?
Real workers and verified experts, not actors. Camera and screen recordings come from the capture programs: tradespeople, cooks, caregivers, office workers and others who are paid per accepted hour to record the work they already do. Expert demonstrations come from the specialists in the open opportunities.
Can we specify exactly what gets recorded?
Yes. Every engagement starts with a written capture protocol covering the tasks, the viewpoint, the variations and what counts as a complete episode. The first sessions are checked against it before anyone records at volume, and you see samples before committing.
What equipment do you use?
For egocentric capture we ship a small head- or chest-mounted camera and, where the protocol calls for it, a depth sensor. Screen capture uses our recorder, limited to the apps the worker chooses. Teleoperation runs on your robot hardware or ours, agreed during scoping.
How do you handle people, customers and patients who appear in recordings?
Recordings showing identifiable people are only offered for training with their consent on file. Faces and anything outside the task are masked, and clinical or care settings are only recorded where the facility and everyone involved has agreed in writing.
Can you record inside our own company?
Yes. Teams join as a group: we handle consent for everyone, every recording hour is paid, and the data can be licensed back to you exclusively, or sold on with your permission. If you already hold years of recordings or records, see selling company data.
What do we receive?
Accepted recordings with the labels in your spec, technical metadata, consent references and a signed provenance record per file, delivered as WebDataset shards, Parquet tables or JSON Lines with linked media and a Croissant dataset card.
Tell us what the model needs to see. We'll go record it.
Five minutes to send a brief. A data lead replies within one business day with a draft capture protocol and a first sourcing plan.