AIWiki
Malaysia
Back to all articles
Tools & PlatformsAI coding agentsoftware engineeringopen source

OpenHands

4 min readUpdated October 2026
OpenHands
Type
Autonomous AI software development agent
Developed by
All Hands AI and the OpenHands community
First released
12 March 2024 (as OpenDevin)
Renamed
26 August 2024
Licence
MIT (core); proprietary licence for enterprise components
Related
Devin, Claude Code, Cline, SWE-bench
OpenHands is an open-source platform for autonomous AI software development. An OpenHands agent takes a natural-language task, opens an isolated sandbox, and then edits files, runs shell commands, executes tests and browses the web until the task is done — as a human developer would, rather than as an inline autocomplete or chat assistant. The project was launched on 12 March 2024 under the name OpenDevin, renamed OpenHands on 26 August 2024, is maintained principally by the company All Hands AI, and is released under the MIT licence.[1][2]

History

OpenDevin was created as a community-built, open alternative to Devin, the proprietary autonomous software engineer introduced by Cognition AI. Its academic and industry creators — including Graham Neubig, Xingyao Wang and Robert Brennan — went on to found All Hands AI, which raised a US$5 million seed round led by Menlo Ventures in September 2024.[1][3] The project described itself in its ICLR 2025 paper as "a platform for the development of powerful and flexible AI agents that interact with the world in similar ways to those of a human developer: by writing code, interacting with a command line, and browsing the web".[2]

The August 2024 rebrand to OpenHands gave the project an identity independent of Devin and signalled scope beyond cloning it. In October 2025 the GitHub organisation moved from All-Hands-AI to OpenHands and documentation moved to docs.openhands.dev. All Hands AI released the Software Agent SDK in November 2025, reached version 1.0.0 on 16 December 2025 with a rewritten, well-typed architecture replacing the original event-stream design, and launched OpenHands Enterprise in May 2026 around an "Agent Control Plane" for managing fleets of coding agents.[1][3]

Architecture and capabilities

OpenHands separates three concerns: the agent, which decides what to do next; the runtime, which executes actions inside an isolated Docker sandbox; and the event stream, an ordered log of every action and observation, which makes sessions auditable and replayable.[1][4] Its default agent, CodeActAgent, implements the CodeAct framework, which unifies the action space around executable code rather than structured tool calls — a single instruction can trigger edits, terminal commands, test runs and web lookups.[2]

The platform is model-agnostic: LLM calls are routed through LiteLLM, so it supports Anthropic, OpenAI, Google, Mistral and Cohere models, plus local deployments via Ollama, vLLM or SGLang, and gateways such as OpenRouter.[1][3] On the SWE-bench Verified benchmark of real GitHub issues, OpenHands has reported resolving 60.6% of issues in a single attempt and 66.4% with a multi-attempt scaling strategy, placing it among the strongest open-source coding agents.[5] Interfaces include a browser-based canvas (Agent Canvas), a command-line client and the SDK; it can run locally, on a remote server or as a hosted cloud service.[3][6]

>See Also

🇲🇾Malaysian Context

🇲🇾 OpenHands matters to Malaysia primarily as a self-hostable, bring-your-own-model alternative to subscription coding agents. Malaysian software houses, agencies and product teams can run the MIT-licensed stack on their own infrastructure, keeping source code and client repositories inside their own environment — an important consideration when contracts require compliance with the Personal Data Protection Act (PDPA) and client data-residency clauses, and where sending code to a third-party cloud is not permitted.

Cost is a second factor: because the agent is model-agnostic, teams can pair it with lower-priced or locally hosted models rather than paying per seat for closed agents, which suits Malaysian SMEs and startups operating on thin tooling budgets. Nationally, the platform supports the talent agenda set out in the National AI Roadmap and the National AI Office's implementation agenda — MDEC's digital skills programmes, HRD Corp-funded training and university computing curricula can teach agentic software development without licensing proprietary tooling. Malaysia's large pool of software engineering talent, concentrated in the Klang Valley, Penang and Johor, is also the pool from which contributors to — and users of — an open agent ecosystem are drawn.[7]

References

  1. ↑Wikipedia. (2026). OpenHands. https://en.wikipedia.org/wiki/OpenHands
  2. ↑Neubig, G. et al. (2024). OpenDevin: An Open Platform for AI Software Developers. arXiv:2407.16741. https://arxiv.org/abs/2407.16741
  3. ↑OpenHands. Official project site and documentation. https://www.openhands.dev
  4. ↑All Hands AI. (2025). The OpenHands Software Agent SDK. https://github.com/OpenHands/software-agent-sdk
  5. ↑OpenAI-independent benchmark reporting. (2026). OpenHands on SWE-bench Verified. https://aiwiki.ai/wiki/openhands
  6. ↑jacar.es. (2026). OpenHands: An Autonomous Coding Agent. https://jacar.es/en/openhands-autonomous-coding-agent
  7. ↑MDEC. Malaysia Digital and AI talent programmes. https://mdec.my
Back to all articles