tools-only Overview
OpenHands is an open-source, model-agnostic platform for building and deploying AI coding agents. It provides a composable Python SDK, CLI, and cloud infrastructure for running agents that write software - from simple tasks like README generation to complex multi-agent refactors. The platform achieves state-of-the-art performance on SWE-bench Verified (77.6%) and serves as the preferred evaluation harness for LLM coding benchmarks.
Core Value Proposition: Scale from a single task to thousands of parallel agents with full control over model selection, deployment, and security.
Problem Addressed
| Problem | How OpenHands Solves It |
|---|---|
| Proprietary AI coding tools lock you into specific models | Model-agnostic architecture works with Claude, GPT, Qwen, Devstral, or any LLM |
| Scaling from laptop to production requires re-architecture | Single Python API runs locally or scales to 1000s of agents in the cloud |
| AI agents lack access to real development environments | Secure, sandboxed Docker/Kubernetes runtime with file system, shell, and browser |
| Evaluating and improving AI coding agents is slow and expensive | 30x speedup on SWE-bench evaluation with cloud-based infrastructure |
| Integrating AI agents into existing workflows requires custom tooling | Native integrations with GitHub, GitLab, Slack, Jira, Linear, and CI/CD |
Key Statistics (as of January 26, 2026)
| Metric | Value |
|---|---|
| GitHub Stars | 67,108 |
| Forks | 8,347 |
| Open Issues | 252 |
| Contributors | 30+ |
| SWE-Bench Verified Score | 77.6% |
| SWE-Bench Lite Score | 41.7% |
| Languages Supported | Python (primary), Any (via Docker) |
| License | MIT (core), Source-Available (enterprise) |
| Created | March 13, 2024 |
Product Offerings
1. Software Agent SDK
The core Python library powering all OpenHands products.
Features:
- Unified Python API for local and cloud execution
- Pre-defined tools for Bash, file editing, web browsing, MCP
- REST-based Agent Server for Docker/Kubernetes deployment
- Custom agent definition and tool creation
- Automatic context compression
- Task planning and decomposition
Usage:
from openhands import Agent, Tool
# Define custom agent
agent = Agent(
model="claude-3-5-sonnet",
tools=[Tool.bash, Tool.file_edit, Tool.web_browse]
)
# Run on code task
result = agent.run("Fix the authentication bug in login.py")
2. OpenHands CLI
Terminal-based interface similar to Claude Code or Codex.
Installation:
pip install openhands-cli
openhands --model claude-3-5-sonnet
Features:
- Works with any LLM (Claude, GPT, open-source models)
- Familiar terminal experience
- Full file system access within sandbox
3. OpenHands Local GUI
Desktop application with REST API and React frontend.
Features:
- Similar experience to Devin or Jules
- REST API for integration
- Single-page React application
- Local laptop execution
4. OpenHands Cloud
Hosted deployment with enterprise features.
Features:
- Free $10 credit to start
- GitHub/GitLab/Bitbucket integrations
- Slack, Jira, Linear integrations
- Multi-user support with RBAC
- Conversation sharing and collaboration
- Usage reporting and budgeting
5. OpenHands Enterprise
Self-hosted deployment for VPC/Kubernetes environments.
Features:
- Source-available (enterprise/ directory)
- Extended support and research team access
- Works with CLI and SDK
- Full control over data and infrastructure
Technical Architecture
State-of-the-Art Agent: CodeAct 2.1
OpenHands includes CodeAct 2.1, an open-source agent achieving top benchmark performance:
Key Innovations:
- Inference-Time Scaling: Multiple solution attempts with critic model selection
- Automatic Context Compression: Manages long-running tasks
- Security Analysis: Built-in code safety checks
- Strong Agent-Computer Interfaces: Reliable tool execution
OpenHands LM 32B: Open coding model (available on Hugging Face):
- 32B parameters (runs on single 3090 GPU)
- 37.2% on SWE-Bench Verified
- No API calls required
Installation & Usage
SDK Installation
pip install openhands-sdk
CLI Installation
pip install openhands-cli
Local GUI (Docker)
docker pull openhands/openhands:latest
docker run -p 3000:3000 openhands/openhands
Cloud Signup
- Visit https://app.all-hands.dev
- Sign in with GitHub or GitLab
- Get free $10 credit
Benchmark Performance
| Benchmark | Score | Notes |
|---|---|---|
| SWE-Bench Verified | 77.6% | State-of-the-art (Jan 2026) |
| SWE-Bench Lite | 41.7% | Open-source leader |
| SWE-Bench Full | Top tier | 2,294 instances |
| SWE-Bench Multimodal | Supported | 517 instances with visuals |
Evaluation Infrastructure: https://github.com/OpenHands/benchmarks
Community & Ecosystem
Related Projects
- Software Agent SDK: https://github.com/OpenHands/software-agent-sdk
- OpenHands CLI: https://github.com/OpenHands/OpenHands-CLI
- Chrome Extension: https://github.com/OpenHands/openhands-chrome-extension
- Theory-of-Mind Module: https://github.com/OpenHands/ToM-SWE
- Benchmarks: https://github.com/OpenHands/benchmarks
Community Resources
- Slack: https://openhands.dev/joinslack
- Product Roadmap: https://github.com/orgs/openhands/projects/1
- Research Paper: https://arxiv.org/abs/2511.03690
Relevance to Claude Code Development
Direct Applications
- Alternative Agent Framework: OpenHands SDK provides patterns for agent composition that could inform Claude Code plugin development
- Evaluation Infrastructure: SWE-bench harness could evaluate Claude Code agent performance
- Tool Design Patterns: Pre-built tools for file editing, shell, and web browsing demonstrate effective tool interfaces
- Model-Agnostic Architecture: Patterns for supporting multiple LLM backends
Patterns Worth Adopting
- Workspace Isolation: Docker/Kubernetes sandboxing for safe code execution
- CodeAct Interface: Combining code and natural language in single turns
- Inference-Time Scaling: Multiple attempts with critic model selection
- Context Compression: Automatic management of long conversations
- REST-Based Agent Server: Scalable deployment pattern
Integration Opportunities
- MCP Integration: OpenHands supports MCP tools - potential for shared tool ecosystem
- Benchmark Suite: Could evaluate Claude Code agents using OpenHands infrastructure
- Open Model Testing: OpenHands LM 32B provides baseline for local model performance
- Enterprise Patterns: Source-available enterprise features provide deployment blueprints
Competitive Analysis
| Feature | OpenHands | Claude Code |
|---|---|---|
| Model Lock-in | None (any LLM) | Claude only |
| Open Source | Yes (MIT core) | No |
| Cloud Offering | Yes | Yes (via Anthropic) |
| Self-Hosted Enterprise | Yes | No |
| SWE-Bench Performance | 77.6% | Not published |
| MCP Support | Yes | Native |
| Local Model Support | Yes (OpenHands LM) | No |
References
- Official Website: https://openhands.dev (accessed 2026-01-26)
- GitHub Repository: https://github.com/OpenHands/OpenHands (accessed 2026-01-26)
- SDK Documentation: https://docs.openhands.dev/sdk (accessed 2026-01-26)
- SWE-Bench Leaderboard: https://www.swebench.com/ (accessed 2026-01-26)
- CodeAct 2.1 Announcement: https://openhands.dev/blog/openhands-codeact-21-an-open-state-of-the-art-software-development-agent (accessed 2026-01-26)
- OpenHands LM 32B Announcement: https://openhands.dev/blog/introducing-openhands-lm-32b-a-strong-open-coding-agent-model (accessed 2026-01-26)
- Research Paper: https://arxiv.org/abs/2511.03690 (accessed 2026-01-26)