OpenHands

OpenHands is an open-source, model-agnostic platform for building and running autonomous AI coding agents. Formerly known as OpenDevin, it provides a Software Agent SDK, CLI, web GUI, and hosted cloud service for software engineering tasks like writing code, fixing bugs, reviewing pull requests, and migrating legacy codebases.

OpenHands is an open-source platform for software-engineering agents, distributed as: a local self-hosted stack (GUI, CLI, and desktop 'Agent Canvas' app), a Python 'Software Agent SDK' for building custom agents, and a hosted cloud version at app.openhands.dev. The core repository (github.com/OpenHands/OpenHands) is MIT-licensed, confirmed both via GitHub's API and the repository's LICENSE file. It began life in March 2024 as 'OpenDevin' — an open response to Cognition's closed-source Devin — and was rebranded to OpenHands under the company All Hands AI later that year.

As of August 21, 2026, the main repo has 84,621 GitHub stars, 11,025 forks, and 519 open issues, with the most recent push the same day this was checked. Releases are frequent: v1.10.0 through v1.14.0 shipped within a 12-day span (Aug 5–17, 2026), each with named contributors, both features and bug fixes, and full changelogs. A pull request was merged the day before this check. This is a genuinely maintained project, not an abandoned scaffold with vendor marketing on top.

Three tiers, per the vendor's pricing page: (1) Open Source/Local — free, self-hosted, full feature set, community support. (2) Individual (SaaS) — free to use with your own API key, or pay OpenHands' hosted model providers at cost with no markup; capped at 10 conversations per day per user. (3) Enterprise — custom pricing, self-hosted in your VPC via Kubernetes or SaaS, SAML/SSO, unlimited concurrent conversations, a 'Large Codebase SDK,' and named support. The vendor's own enterprise page describes that self-hosted tier as 'source-available' rather than open source — distinct language from the MIT core repo — so the code your security team can 'inspect and audit' at that tier may not carry the same redistribution rights as the free OSS product.

OpenHands is built by All Hands AI, Inc., co-founded by Graham Neubig, Robert Brennan, and Xingyao Wang. The company raised a $5M seed round in September 2024 (Menlo Ventures leading, with Pillar VC, Betaworks, Rebellion, and angel investors including Hugging Face co-founder Thom Wolf), followed by an $18.8M Series A in November 2025 led by Madrona, with Menlo Ventures, Pillar VC, Obvious Ventures, Fujitsu Ventures, and Alumni Ventures participating — $23.8M disclosed total. The Series A announcement also disclosed a strategic partnership with AMD.

A high-severity command-injection vulnerability (CVE-2026-33718, CVSS 7.6) was found in OpenHands' git-diff handler — an authenticated user could pass shell metacharacters through a path parameter to execute arbitrary commands in the agent's sandbox. It was reported by external researchers, reviewed and published as a GitHub Security Advisory, and fixed in v1.5.0. That's the disclosure process working as intended, though it's also a reminder that a tool designed to execute shell commands on your behalf carries that class of risk by nature — worth factoring into how you sandbox any self-hosted deployment. Separately, the vendor runs a Vanta-powered Trust Center (app.vanta.com/openhands.dev/trust) that publishes live-monitored security controls across infrastructure, organizational, product, and data-privacy categories, plus a subprocessor list (Google Cloud Platform, Datadog, and Anthropic for the hosted product's AI models). We could not confirm a completed, published SOC 2 report from the public trust center pages — the vendor lists continuously-monitored controls but a formal audit certificate wasn't visible without an access request.

The project has had multiple Show HN and follow-up Hacker News threads since its 2024 launch (covering the SDK, an AMD partnership, and the Series A), indicating sustained community engagement, though we did not verify detailed comment sentiment. There's no presence on G2, Capterra, or Trustpilot — unsurprising for a developer tool distributed primarily through GitHub rather than a review-driven SaaS funnel; treat that absence as normal rather than a red flag.

• Software Agent SDK: A composable Python library for defining and running agents locally or at scale, with model-agnostic orchestration through the litellm library
• Sandboxed Execution: All code runs inside secure Docker containers, so agents can write, test, and iterate on code without risking your local environment
• Model-Agnostic Support: Works with OpenAI, Anthropic Claude, Google Gemini, local models through Ollama or vLLM, and 75+ other providers through litellm
• GitHub and GitLab Integration: Agents clone repositories, create branches, push commits, and open pull requests directly within existing repo workflows
• Local GUI and CLI: A browser-based React interface with embedded VS Code editor, terminal access, and a separate CLI for terminal-first workflows
• OpenHands Cloud: Hosted deployment at app.all-hands.dev with Slack, Jira, and Linear integrations, plus a free tier with 10 daily conversations
• Skills System: Specialized prompts with domain-specific knowledge that let teams adapt agents to particular workflows without rebuilding core logic
• SWE-Bench Performance: Resolves 53%+ of real-world GitHub issues on SWE-bench Verified when paired with strong models, scoring 77.6 on the benchmark

• Engineering teams handling maintenance backlogs: Use OpenHands to automate repetitive tasks like fixing merge conflicts, resolving linter failures, refactoring modules, and triaging GitHub issues across multiple repositories
• Backend developers migrating legacy codebases: Run agents to navigate large codebases, convert code between languages (COBOL to Java, for example), and test changes across multiple modules with limited manual work
• DevOps teams running automated security workflows: Set up agents for vulnerability scanning that automatically generate pull requests with fixes, integrated into existing CI/CD pipelines
• Solo developers prototyping applications: Use the Planning Agent to scope features from a high-level prompt, review a generated plan, then let the agent handle implementation, testing, and debugging
• Not ideal for: Non-technical users who need point-and-click automation (use Make, Zapier, or n8n instead) or teams that need pre-tuned domain-specific agents with strict compliance requirements

Strengths:

• Fully open-source with MIT licensing, so teams can self-host, audit, and modify without vendor lock-in
• Model-agnostic design means you can swap between cloud providers or run local models based on cost and privacy needs
• Sandboxed Docker execution keeps agent actions isolated and auditable, a real advantage for enterprise security requirements
• The free open-source tier has no conversation limits, and the cloud individual tier starts at $0 with pay-as-you-go LLM usage at cost

Weaknesses:

• Not fully autonomous yet, as users report needing to provide corrections and guidance during longer tasks
• Sandbox setup can be tricky, with reports of crashes and outdated documentation after the v0.10 release
• Rapid release cadence has introduced breaking changes in agent protocols, which can frustrate teams on older versions
• Error messages are sometimes vague, including reports of tasks failing without useful tracebacks
• The free cloud tier is limited to 1 workspace and 10 daily conversations

Is OpenHands free?

Yes. The open-source version is completely free with no conversation limits. The cloud individual tier is also free with 10 daily conversations and pay-as-you-go LLM costs at provider rates with no markup.

What was OpenHands previously called?

OpenHands was previously called OpenDevin.

Does OpenHands support local models?

Yes. OpenHands works with local models through Ollama, vLLM, and any provider supported by litellm. You can run it entirely on your own infrastructure with no external API calls.

How does OpenHands compare to Cursor?

OpenHands handles autonomous, long-running software engineering tasks in a sandboxed environment with support for 75+ LLM providers. Cursor is better for IDE-based coding with deep codebase indexing and real-time suggestions. Choose OpenHands for agent-driven automation; choose Cursor for daily editing.

Can OpenHands run code safely?

Yes. All code execution happens inside secure Docker containers. The sandbox isolates agent actions from your host system, and enterprise deployments can run in private VPCs with full RBAC controls.

How do I install OpenHands?

Clone the GitHub repo, run make build and make dev with Docker. On first launch, configure your LLM API key and model. Most developers report getting a working setup in 15 to 30 minutes.