Data and evals for long-horizon digital agents
Traceform turns real computer and browser interactions into training-ready trajectories, powering reinforcement learning environments for real-world work.
Powering the next generation of
Everything you need to build
General Agents.
From data collection to deployment, we provide the primitives so you can train, iterate, and measure like it's a real ML system.
Datasets
Large-scale, high-signal UI data: screenshots, DOM trees, actions, and outcomes.
RL Envs
Deterministic, resettable sandboxes with step-level instrumentation.
Initializing Ubuntu 22.04...Connected to VNC:5900Reward hook active.Evals
Replay runs, fork checkpoints, and compare policies against ground truth.
Trusted by
Frontier Labs
Pre-train state-of-the-art VLMs on millions of hours of curated desktop interaction.
Researchers
Benchmarking reasoning & safety.
Enterprises
Secure, private automation.
Data Teams
Scalable collection with full provenance.
How it works
Spin up desktops
Windows/macOS/Linux with instrumented agents
Record tasks
Humans, scripted bots, or agents into trajectories
Package datasets
Clean, label, redact, verify
Train + evaluate
Offline RL / imitation / online RL with replays
The complete stack
for computer-use agents.
Data
Trajectory capture at scale with full state tracking and ground truth labels.
Environments
Resettable desktop sandboxes.
➜ Initializing sandbox environment...➜ Mounting VNC filesystem...ℹ GPU acceleration enabled (RTX 4090)➜ Loading generic-agent-v4...➜ Setting viewport: 1920x1080✔ Ready. Connect at port 5900.> Navigating to kayak.com...> Selecting date: 2024-12-15Training
From demo to SOTA.
Built for diverse industries
Training data that understands compliance requirements across healthcare, finance, and legal workflows.
Start collecting trajectory data today.
Glmpse runs quietly in the background, capturing screenshots, DOM state, and user actions with full privacy controls. Perfect for building computer-use agent training data.