Computer Use
Agent Framework
Run AI agents in isolated Docker containers. Watch them work in real-time. Verify results automatically. Multi-provider support built-in.
$ uv run helios tasks/explore-desktop --watch
# Agent running in Docker container
# Open http://localhost:8080 to watchCapabilities
Everything you need to run
AI agents at scale
A complete framework for orchestrating computer-use agents with real-time observation and automated verification.
Isolated Environments
Each agent runs in its own Docker container with full desktop access. Complete isolation ensures safe execution of any task.
Real-time Viewing
Watch agents work through a live web viewer at localhost:8080. See every click, keystroke, and decision as it happens.
Automated Verification
Define test.sh scripts to verify agent outcomes. Get clear pass/fail results with granular reward scores from 0 to 1.
Multi-Provider
Switch between Anthropic, OpenAI, Gemini, and AWS Bedrock with a single flag. Use the best model for each task.
Batch Execution
Run multiple tasks in parallel with configurable concurrency. Perfect for benchmarks and large-scale evaluation runs.
Cloud Ready
Deploy to Daytona cloud sandboxes for scalable execution without local Docker. Enterprise-ready from day one.
How it works
Simple architecture,
powerful execution
Helios orchestrates a clean pipeline from task definition to verified results. Every component is modular and extensible.
Task
Define what the agent should do with instruction.md and task.toml configuration
Gateway
Route to any LLM provider through a unified, type-safe interface
Environment
Execute in isolated Docker containers or scalable cloud sandboxes
Verifier
Run test.sh scripts and collect reward scores automatically
reward.txt0 | 1 | 0.0-1.0Integrations
Use any major LLM provider
$ helios tasks/my-task -m claude-sonnet-4-20250514Quick Start
Get started in 60 seconds
Install, configure, and run your first agent.
# Install dependencies
$ uv sync# Set up API keys and build Docker images
$ cp .env.example .env
$ docker build -t cua-desktop -f docker/Dockerfile.desktop .# Run your first agent with live viewing
$ uv run helios tasks/explore-desktop --watch
# Open http://localhost:8080 to watch the agent
Ready to build with
AI agents?
Join developers using Helios to run, observe, and verify computer-use agents at scale.