Skip to content

The best session capture agent for computer-use training — works on real software.

Where AI learns to work.

Capture real expert work as structured training data, and let agents practice in real computer environments - imitation and practice, one platform.

Book a demo >Request beta access >
Why now

The frontier has moved
from exams to occupations.

AI is no longer measured on benchmarks — it's measured on real files, real deliverables, real professional work. What models could do in principle keeps racing ahead of what they execute reliably.

THAT DISTANCE IS THE TRAINING GAP — A DEMAND CURVE, NOT A FINISH LINE.

Layered panes: real work, capture, environments — the training stack

THE TRAINING STACK● REAL WORK, MADE PRACTICABLE

How models learn

Three stages of training. The last two depend on real work.

01Text → patterns

Pretraining

Trillions of tokens of text teach language, facts, and reasoning patterns. Knowledge — not competence.

02Demonstrations → behavior

Supervised fine-tuning

Experts perform real work, recorded step by step. The model trains on the demonstrations and reproduces the behavior.

03Outcomes → reliability

Reinforcement learning

The model attempts the task itself, gets scored on the outcome, and updates. Every run gets further than the last.

Capture Suite · 01

Imitation needs recorded expert sessions.

Capture Agent >

Environment Suite · 02

Practice needs scalable training environments.

RL Environments >
Platform

One platform. From expert hands to trained agents.

Suite 01>

Capture Suite

Turn expert sessions into structured training assets — screen, input, voice, files.

Suite 02>

Environment Suite

Boot real computers as RL environments and run them at fleet scale.

Suite 03>

On-Prem Fine-Tuning

Capture inside your walls and train corporate agents on your own work data.

Why PixelNode

Two ways to get training data. Only one survives contact with real work.

Everything needed to teach models real work.

⚠ STATUS: UNRESOLVED

Forcing it with
“almost-real” data

The synthetic set plateaued — eval's been flat for three weeks.
Vendor sent 40 hours of screen video. No events, no files, no timestamps.
Sandbox broke again after the app auto-updated. Rebuilding.
Who verified these 12k outputs? — Nobody.
✕ ENV DRIFT — restore mismatch● 2,931 SESSIONS UNLABELED✕ LICENSE EXPIRED — cad-env-04⚠ REWARD HACKING DETECTED● QA BACKLOG — 3 WEEKS

Training with

Pixelnode
✓1,204 expert sessions captured today
The data wall

The model needs the work, not a description of it.

Synthetic sandboxes create volume. Pixelnode captures the actions, reasoning, machine state, and verified outcome of real expert work.

Before
Generated examples

Synthetic volume

Generated examples without a real application state, expert intent, or verified outcome.

After
Captured trajectory
SOLIDWORKS / robot_joint_A1.sldasmREC  00:47:12
XYEXPLODED · ISO
00:47:12component_selectedharmonic_drive / output_flange
VOICESelecting the output flange before defining the mate.
SELECTMATESTATEVERIFIED

Recorded expertise

An expert's actions, reasoning, files, machine state, and verified outcome, captured as one training trajectory.

Capture Suite · 01

Imitation needs recorded expert sessions.

The capture agent turns real expert work - screen, input, voice, files, machine state - into structured, replayable training data.

See how capture works >
Capture Suite · 01

One expert hour becomes a structured training asset.

PROMPT──SCREEN + INPUT RECORDING──DELIVERABLE MONITORING──LOGS + THINK-ALOUD──PRIVACY MODE──>MODEL TRAINING DATA
Prompt

A task brief goes in. A complete, structured, replayable record of expert work comes out — synchronized to the frame, ready for training.

Everything we capture >
capture.pixelnode.ai/session/expert_014REC 00:47:12
SCREENINPUTMICFILESPRIVACY AUTO
cut_v2.prprojassets/notes.md
Capture Suite · 01

Beyond screen capture and keystrokes, every interaction becomes a data point.

session/expert_014 · signalsREC
00:46:02 app_focus — Premiere Pro
00:46:18 tool_used — lumetri.color.wheels
00:46:29 shortcut — ⌘⇧E export dialog
00:46:41 ui_click — timeline.track_02
00:46:55 tool_used — razor @ 00:12:04:11
Capture Suite · 01

Desktop · live capture

REC
WORK — EVERY FRAMEPERSONAL — 0 FRAMES

Privacy is a capture feature, not an afterthought.

  • 1.1

    Computer Permission Management

    Explicit, revocable permissions per capture — the expert stays in control.

  • 1.2

    Window-Level Privacy Mode

    Exclude any window from the record, surgically — the work stays, the personal goes.

  • 1.3

    On-Demand & Automated Privacy

    Pause by hand, or let detection pause for you when something private appears.

Capture Suite · 01

The deliverable trail, from first save to final export.

file_created

cut_v1.prproj

file_saved

cut_v1.prproj

SNAPSHOT

restore_point_09

file_renamed

cut_v2.prproj

SNAPSHOT

restore_point_10

file_exported

final_v2.mp4

RESTORING MACHINE STATE → restore_point_09 · ← restore the whole machine to any captured state

File Events Tracking

Creates, saves, renames, exports — every deliverable event, timestamped.

File / Folder Snapshots

Point-in-time copies of the working set, tied to the session timeline.

Computer Restore Points

Return the whole machine — apps, files, state — to any captured moment.

Capture Suite · 01

Built for real expert setups.

Whatever the craft demands — the capture agent records it natively.

01 / 05

Dual Monitor

Full multi-display capture — reference on one screen, work on the other, both recorded as one session.

02 / 05

Audio & Microphone

System audio and voice narration on separate tracks, sample-synced to every frame of the recording.

03 / 05

Drawing Tablet + Pen

Full stylus telemetry — pressure, tilt, and hover — captured alongside the canvas it shapes.

04 / 05

Touch Input

Native touch events from gestures to multi-finger interactions, recorded as structured input data.

05 / 05

Webcam

An optional context track — what the expert sees and reacts to, kept private and opt-in by default.

Capture Suite · 01

Keep the expert authenticity — capture from any device.

Cloud workstationsPROVISIONED 000/100

For scalability

Provisioned desktops with apps preinstalled — spin up a hundred identical capture stations before lunch.

Local devicesREC

For authenticity

Experts work on their own machines, their own setups, their own muscle memory — the data stays honest.

Windows

Native capture agent

v2.4 STABLE

macOS

Native capture agent

v2.4 STABLE

Linux

Native capture agent

v2.4 STABLE
Environment Suite · 02

Practice needs scalable training environments.

After imitation comes practice - agents attempt real tasks on real computers, get scored on the outcome, and improve run over run.

Explore environments >
Environment Suite · 02

Everything practice needs, built into every environment.

Ready-to-use environments

Start real working environments instantly, with the tools and files each occupation needs.

Native computer use

Models work through the same display, pointer and keyboard signals people use.

Automatic snapshots

Capture the machine, open apps and working files at useful checkpoints.

Restore from any point

Restart practice from a chosen state without rebuilding the environment.

Environment Suite · 02

Repeatable environments. Verifiable results.

2.51 checkpoint → 1,000 runs

Checkpoint forking

Freeze minute 47 of a real expert session and let agents practice recovering — a thousand times.

2.6Checks → verdict

Result Verifier API

The automatic referee. Programmable checks against files, app state, and exports.

2.7Compose → version

Prebaked golden images

Versioned machine states. Compose a "level" per occupation in minutes.

2.8reset · step · observe

Agent gateway

The same channel a human uses. Adapters for every major trainer.

2.91.8s · 18 regions

Instant restore & fleet

Byte-identical restore in seconds, not VM-minutes. Parallel episodes, isolation, metering — across all major clouds and regions.

Environment Suite · 02

Ready-to-use environments, composed per occupation.

214 GOLDEN IMAGES
58 APPLICATIONS · ALL REGIONS

Icon render

Built for every kind of work.

From creative work and engineering to finance and legal, any desktop workflow can become a repeatable training environment.

  • Adobe Premiere Pro
  • Microsoft Excel
  • Blender

+ any desktop application

See the live fleet
Fleet

Any occupation. Any application. Infinite practice runs.

LIVE FLEET
18,204 EPISODES TODAY

1,024 active environments18 regions41ms median gateway99.97% restore integrity
On-Prem Fine-Tuning · 03ENTERPRISE PRIVATE BETA

Your experts. Your data. Your agents. The same stack — inside your walls.

3.1

Local Capture

Deploy capture on company workspaces and record company-specific workflows.

3.2

On-Premise Data Storage

Capture data lives in your own data centers. It never leaves the building.

3.3

Custom Trained Agents

Agents do company work — trained on your own company data.

Talk to us about on-prem >
Who it's for

Built for the teams building working AI.

Frontier AI labs

01

Human-data infrastructure as a service.

  • Expert capture operations at scale — cloud workstations or local devices
  • Custom occupation environments from prebaked golden images
  • Agent gateway adapters for your training stack
  • Verified, structured data pipelines — session to dataset
Talk to us >

Talent platforms

02

White-label the stack. Keep your brand.

  • Capture Suite under your own brand, for your own workforce
  • Onboarding, QA, and structuring tooling for sessions at scale
  • Access to our 250K+ professional network for recruiting
  • Co-sell environments alongside your data offerings
Talk to us >

Enterprises

03

Your workflows, your data, your agents.

  • Local capture on company workspaces
  • On-premise data storage — nothing leaves the building
  • Custom agents trained on your own work data
  • Consent, permissioning, and privacy controls built in
Talk to us >
Built consent-first, from pixel to dataset.Consent-first capturePermission managementEnvironment isolationWindow-level privacyOn-prem residencyAutomated privacy modes

Vision

Every company will train their own agents.

PixelNode turns expert workflows, judgment and institutional knowledge into private training data for agents that learn how your company actually works.

CapturedStoredTrained In-House

Book now
Training loopCapture
Capture

Expert work is recorded across screen, input, voice and files.

Security architecture

Boundary intact

A transparent industrial lock revealing its mechanical security core
  • ISO 27001
  • SOC 2
  • GDPR

Your training datastays inside the boundary.

Security is part of the training architecture, not a layer added later. PixelNode protects captured work from the first frame through the final training run.

  1. 01

    Encrypted by default

    Capture data is encrypted in transit and at rest.

  2. 02

    Controlled access

    Role-based access and hardware-backed authentication protect raw sessions.

  3. 03

    Every access recorded

    Raw-session access is written to an audit log.

  4. 04

    An on-prem boundary

    Capture, redaction, storage and fine-tuning can stay inside your infrastructure.

Read our security policy

Private beta — now onboarding design partners

See the training
layer live.

Book a demo >