Openhands

title: "OpenHands" description: "Autonomous AI software engineering agent (formerly OpenDevin) that can write code, run commands, browse the web, and interact with files. Runs in sandboxed Docker environments. Supports 50+ LLM backends. Use for automated research coding, experiment implementation, and autonomous repository management." skillName: "openhands" skillVersion: "1.0.0" skillAuthor: "Orchestra Research" skillLicense: "MIT" skillTags: ["Agents", "OpenHands", "Autonomous Coding", "Software Engineering", "Docker", "Code Generation", "Research Automation", "Multi-LLM"] skillDeps: ["openhands-ai", "docker"]

OpenHands — Autonomous AI Software Engineer

Formerly OpenDevin. An AI agent that writes code, runs experiments, and manages repositories autonomously.

When to use OpenHands

Use OpenHands when:

• Automating research experiment implementation
• Autonomous codebase exploration and modification
• Running multi-step coding tasks without human intervention
• Integrating with GitHub for PR creation and review
• Building research pipelines that write their own code

Metrics:

• 40,000+ GitHub stars
• 50+ LLM backends supported
• Sandboxed Docker execution (safe by default)
• Can browse web, run terminal, edit files autonomously

Quick start

# Using Docker (recommended)
docker pull docker.all-hands.dev/all-hands-ai/runtime:0.38-nikolaik

docker run -it --rm --pull=always \
    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.38-nikolaik \
    -e LOG_ALL_EVENTS=true \
    -v /var/run/docker.sock:/var/run/docker.sock \
    -p 3000:3000 \
    --add-host host.docker.internal:host-gateway \
    --name openhands-app \
    docker.all-hands.dev/all-hands-ai/openhands:0.38

# Access web UI at http://localhost:3000

Python API

from openhands.core.main import create_runtime, run_controller
from openhands.core.config import AppConfig, SandboxConfig, LLMConfig

# Configure
config = AppConfig(
    llm=LLMConfig(
        model="claude-sonnet-4-5-20250929",
        api_key="your-api-key"
    ),
    sandbox=SandboxConfig(
        base_container_image="python:3.12-slim"
    )
)

# Run autonomous task
task = """
Implement a GRPO training loop using TRL.
The script should:
1. Load a base model (e.g., Qwen2.5-1.5B)
2. Define a reward function based on format compliance
3. Run 100 training steps and log metrics to W&B
4. Save the final checkpoint
"""

await run_controller(config=config, initial_user_action=task)

Research automation workflow

from openhands_ai import OpenHands

client = OpenHands(api_key="your-key")

# Autonomous experiment implementation
session = client.create_session(
    llm_config={"model": "claude-opus-4-5", "temperature": 0.1},
    sandbox_config={"image": "pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime"}
)

result = session.run("""
Read the paper at ./papers/flash_attention_3.pdf.
Implement the core algorithm in PyTorch.
Run a benchmark comparing throughput vs standard attention on seq_len=[512, 1024, 2048, 4096].
Generate a plot and save to ./results/benchmark.png.
""")

print(result.summary)
print(result.files_created)

GitHub integration

# Resolve GitHub issues autonomously
resolver = GitHubIssueResolver(
    github_token="ghp_...",
    llm_config={"model": "claude-sonnet-4-5-20250929"},
)

# Process a batch of issues
resolver.resolve_issues(
    repo="your-org/research-repo",
    issue_numbers=[42, 43, 44],
    max_iterations=50  # per issue
)

Key capabilities

Supported LLM backends

# config.toml
[llm]
model = "claude-sonnet-4-5-20250929"
# OR
model = "gpt-4o"
model = "gemini/gemini-2.0-flash"
model = "ollama/llama3.2"
model = "deepseek/deepseek-chat"

Best practices for research

• Use Claude Opus/Sonnet for complex multi-step research tasks
• Set max_iterations=100 for long experiments
• Mount your dataset directory as a volume: -v /data:/workspace/data
• Use LOG_ALL_EVENTS=true to trace agent decisions
• For reproducibility: pin the runtime container version

Common pitfalls

• Docker not running: OpenHands requires Docker daemon
• Context window exceeded: Break large tasks into subtasks
• Infinite loops: Set max_iterations as a hard stop
• Slow startup: Runtime container pull can take 2–5 minutes first time

References

• OpenHands GitHub
• Documentation
• Model Support