Frontier labs & model builders who need reliable evals + RL training infrastructure and high-quality task datasets. Teams building RL environments / task datasets to sell to frontier labs. Teams who want to evaluate frontier model capabilities.

Does it work with any model?

Yes. Any model can run against any environment, and HUD wires up the best available tools for each pairing, so Claude, GPT, Gemini, Grok, and your own trained checkpoints all hit state-of-the-art behavior on the same harness, through one OpenAI-compatible endpoint.

Can I train a model on my own environments?

Yes. Fork a trainable base model, point it at one of your tasksets, and HUD trains a new checkpoint; Tinker is our primary training backend. Use it through the same gateway, same interface, better performance on your tasks.

What are HUD QA Agents?

Analysis agents that run automatically on every trace to tell you whether your evals are measuring what you think. Five standard agents — Failure Analysis, Prompt-Grader Alignment, False Negative Detector, False Positive Detector, and Reward Hacking Detector — surface root causes, grader drift, unjustified failures, brittle passes, and shortcut-taking so your reward signal stays honest.

Can I monetize environments I build?

Yes, through the HUD vendor marketplace at https://datavendor.ai/. Publish your environments and tasksets, and get paid when frontier labs source your post-training data. HUD handles the buyer/seller side, entitlements, and trace-level QA so what gets delivered meets spec.

Is HUD open source? How do I get started?

Yes, HUD is open source under the MIT license. You can find the source code on our GitHub repository at https://github.com/hud-evals/hud-python. To get started, install the SDK with uv tool install hud, then hud login. Scaffold an environment with hud init my-env and run it locally with hud dev.

HUD — Build RL environments

The platform for building RL environments

Encode your expertise into environments to train and evaluate models, and create the post-training data that aligns AI to your work.

4,831,034

Task runs

13,764

Environments built

22,297,355

Inference calls

Infrastructure for aligning AI to the real world.

The previous generation built apps to impact the world.This one builds the environments that align AI to it.

Create trainable environments

Start from a community environment, build your own with the SDK, or import existing tasks from Harbor.

Environments

Inspect your traces

Debug agent behavior and improve reward signal by understanding what's actually happening.

Review tasks as a team

Assign tasks, discuss changes in comments, and approve each version as your team iterates.

Turn tasks into revenue

List them on DataVendor, where AI labs buy training and evaluation data.

datavendor.ai

QA your tasks automaticallyAgents audit every trace for grader mistakes and reward hacking.

Failure analysisSurfaces every problem behind a failed run with evidence and a code-level root cause — not just a one-line label.
Prompt-grader alignmentVerifies the grader's checks line up with what the prompt actually asks for, so evals judge the right behaviors.
False negative detectorCatches unjustified failures — the agent actually solved the task, but the grader scored it wrong.
False positive detectorSpots passes that aren't really passes — the agent got credit without genuinely completing the task.
Reward hacking detectorFlags traces where the agent gamed the eval — test manipulation, output hardcoding, scorer exploits.

Failure analysisSurfaces every problem behind a failed run with evidence and a code-level root cause — not just a one-line label.
Prompt-grader alignmentVerifies the grader's checks line up with what the prompt actually asks for, so evals judge the right behaviors.
False negative detectorCatches unjustified failures — the agent actually solved the task, but the grader scored it wrong.
False positive detectorSpots passes that aren't really passes — the agent got credit without genuinely completing the task.
Reward hacking detectorFlags traces where the agent gamed the eval — test manipulation, output hardcoding, scorer exploits.

Customer stories

Sharpe

How Sharpe scales high-quality RL environments

Sharpe designs RL environments for coding, AI training, and ML project management using HUD, with full QA on every task they make to ensure the highest quality training data.

Coming soon

UiPath

How UiPath benchmarks enterprise agents against every frontier model

UiPath brought its UI-CUBE enterprise computer-use benchmark onto HUD, where its own agents and every frontier model are measured against the same enterprise workflows.