The platform for evaluations and RL environments for agents.
Wrap real software as environments, run benchmarks, and train with RL – locally or at scale.

📅 Hop on a call or email us at 📧 founders@hud.ai

HUD is an open-source platform for evaluating and training agents:

• Turn any piece of software into an RL environment.
• Run reproducible benchmarks and submit to public leaderboards.
• Train agents with RL (GRPO) on top of those environments with any RL provider.
• Inspect every tool call, observation, and reward in real time on hud.ai.

The core Python SDK lives in hud-evals/hud-python.

We love contributions!

Check our open hud-python issues and open pull requests for more information.

• 🧱 MCP environment skeleton – wrap any system as an MCP server so any agent can call it.
• 📡 Live telemetry – traces for every trajectory and tool call, streamed to hud.ai.
• 📊 Public benchmarks & leaderboards – OSWorld-Verified, SheetBench-50, and more at hud.ai/leaderboards.
• 🌐 Cloud browsers – integrations for browser-based environments (AnchorBrowser, Steel, BrowserBase, etc.).
• 🔁 Hot-reload dev loop – hud dev for iterating on environments without rebuilds.
• 🧪 One-click RL – hud rl to train GRPO policies on any HUD dataset.