Chakra Labs Research Initiative
High-fidelity trajectories and
environments for frontier AI research
Fig 1: Instructional video engineered for emotional context
Dojo
Collaborative reinforcement learning environments for computer use research.
An open platform for interactive clones of production software built to the same standard as frontier lab infrastructure. Spin up tasks, train and evaluate your models against the environments, and iterate throughout your research. Native support for Verifiers and Verl.
Methodology
Built for Speed
and Scientific Rigor
Chakra environments demonstrate strong generalization across diverse tasks to OSWorld. We've built isolated evaluation protocols that prevent reward hacking during training.
Our infrastructure prioritizes execution speed to enable rapid experimental iteration. All environments are delivered via containerization, with support for both hosted deployments and on-premises installations through Harbor.
Model Performance
50%
40
30
20
10
0
41.39%
29.51%
9.43%
3.69%
Hours of trajectories
Segmented image tokens
Screenshot / action pairs
Automated task generation
Bespoke task generation delivered in 24 hours. Our environments generate novel challenges automatically, enabling experiments with extended autonomous runtime. Models learn to recover from errors and adapt to new situations without human intervention.
GUI + API environments
Environments supporting simultaneous GUI and API interaction within single tasks. Agents can switch between visual navigation and programmatic commands based on task context. This mixed-modality capability enables research on workflows where both interaction modes are available.
Publications
Join us
Chakra is an applied research team pushing the boundaries of agents
We provide research-grade infrastructure for frontier-scale experiments in proprietary regimes where emergent behaviors develop. We work on hard infrastructure problems: frame-accurate state capture, deterministic environment design, mixed-modality training systems.
Realistic interactivity applies to our digital environments only - office space pictured is hypothetical.