OpenHands, formerly known as OpenDevin, is 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. It is introduced in OpenHands: An Open Platform for AI Software Developers as Generalist Agents, which presents it against the background of rapid development in AI agents that interact with and affect change in their surrounding environments.
It is released under the permissive MIT license, and the introducing paper describes it as a community project spanning academia and industry, with more than 2.1K contributions from over 188 contributors at the time of writing. Its code is stated to be at https://github.com/All-Hands-AI/OpenHands.
The project’s first-anniversary post, One Year of OpenHands: A Journey of Open Source AI Development, gives its origin. It began on 12 March 2024 with the aim of creating an open-source AI agent for software development, and was originally created as OpenDevin — an homage to Devin by Cognition, which the post says inspired the project — by Binyuan Hui and Junyang Lin of the team developing the Qwen open-source language model. The open-source community rallied around it, and three early contributors, Graham Neubig, Robert Brennan and Xingyao Wang, went on to found All Hands AI to carry on its mission, at which point the project was renamed OpenHands.
Capabilities
The introducing paper describes four things the platform allows for: the implementation of new agents; safe interaction with sandboxed environments for code execution; coordination between multiple agents; and the incorporation of evaluation benchmarks.
In a March 2026 post, Learning to Verify AI-Generated Code, the OpenHands team describes adding verification to this: a layered “verification stack” whose first layer is a small, fast Critic Model that scores an agent’s whole trajectory, trained on real-world production traces rather than benchmark data. The critic is integrated into the OpenHands Software Agent SDK, where its score can be used in custom loops for reranking, early stopping or iterative refinement, and into the OpenHands CLI, where it can be enabled with early stopping, used for iterative refinement, and given different acceptance thresholds. The post names a second, patch-level layer as still to come.
Context condensation
An April 2025 post, OpenHands Context Condensensation for More Efficient AI Agents, introduces the OpenHands context condenser. Once a conversation grows beyond a threshold, older interactions are summarized — encoding the user’s goals, the agent’s progress and what remains, and for software tasks details such as critical files and failing tests — while recent exchanges are kept intact. On a subset of SWE-bench Verified instances, the post reports per-turn API cost settling at less than half the baseline agent’s, with an average solve rate of 54% against the baseline’s 53%. It describes context condensation as available in OpenHands, including OpenHands Cloud.
Security controls
An August 2025 post from the OpenHands team, Mitigating Prompt Injection Attacks in Software Agents, describes how OpenHands addresses the prompt-injection risk of an agent that typically runs without a human watching it. Every conversation with the agent takes place inside a Docker container, so the local filesystem and environment are kept safe even if the agent runs a destructive command, though the post notes that secrets inside that environment, and the source code, can still be exposed. The OpenHands CLI can ask the user for confirmation before the agent acts, and OpenHands has added a security analyzer, developed with Invariant Labs, that assesses the actions the agent wants to take for potential threats and can stop to ask for explicit permission or refuse to proceed. Organizations that self-host OpenHands with its Kubernetes deployment can additionally restrict the agent inside its sandbox with network policy and eBPF tooling. The post also reports that how readily the agent refuses a malicious instruction depends on the model powering it.
Adoption & Ecosystem
Because evaluation benchmarks are incorporated into the platform itself, the benchmarks it holds are part of what it offers. The introducing paper reports an evaluation of agents over 15 challenging tasks based on the benchmarks currently incorporated, spanning software engineering — for which it names SWE-Bench — and web browsing, for which it names WebArena, among others.
The anniversary post describes OpenHands being used both by individual developers automating routine tasks and by large teams on complex refactoring projects. It set out four aims for the project’s second year: a general release of OpenHands Cloud, then in beta, which runs OpenHands without the user’s own server and can be used from a browser, a mobile device or GitHub; continued improvement of the agent, both in core accuracy on benchmarks such as SWE-Bench and in the range of tasks it performs reliably; more integrations, including a better CLI and task-management tools such as Jira and Linear; and more customization to users’ preferences, organizational policies and workflows.