OpenHands is an open-source platform for autonomous AI software agents, built by All Hands AI. It lets an agent act as a full software engineer: writing code, running terminal commands, browsing documentation, calling APIs and opening pull requests. Given a GitHub issue or a task description, OpenHands plans and executes the whole solution without per-step direction. It has 93.5k GitHub stars and is MIT licensed.
The agent works in an isolated workspace with a terminal, code editor, browser and file system, on a local agent-server where Docker is optional, on a self-hosted VM, or on the managed OpenHands Cloud. It installs dependencies, runs test suites, debugs failures and iterates until the task completes. Verified integrations are GitHub, GitLab and Bitbucket for source control, Jira for issue tracking, Slack for notification and Linear for planning, with Azure DevOps available on the Enterprise tier. MCP support ships on every plan, and it is client-side: OpenHands connects out to external MCP servers and uses their tools automatically, rather than publishing an endpoint other agents connect into. The product now ships as a set of components. Agent Canvas is the local visual workspace and control center, installed on macOS, Linux or Windows, and it can also run Claude Code, Codex and Gemini CLI through the Agent Client Protocol, the Software Agent SDK and Agent Server run the agent, an Automation Server runs scheduled and event-driven automations, and the CLI remains a supported terminal interface. The earlier Docker-based Local GUI is deprecated in favor of Agent Canvas.
All Hands AI publishes the OpenHands Index, a public benchmark that is unusually strong evidence for this category. It scores model and SDK combinations across five categories of real software engineering work: Issue Resolution, Greenfield, Frontend, Testing and Information Gathering. Issue Resolution aggregates SWE-bench and SWE-bench-multimodal. The leaderboard publishes per-category score, dollar cost and runtime for 33 model configurations, 17 open and 16 closed, with downloadable logs for each run.
There are three tiers. Agent Canvas is free and open source under MIT, runs on a laptop or shared VM for a single user, and works with any model through your own key or with other agents such as Claude Code and Codex. Teams is $500 a month for up to 10 users plus $50 a month for each extra seat, self-hosted, with a 30-day trial, multi-user identity and RBAC, SAML and SSO, budgeting controls and built-in integrations, and it is capped at 5,000 sessions a month. Enterprise is custom-priced and adds unlimited seats and users, an enterprise agreement, Standard or Premium support and a customer success manager. The managed OpenHands Cloud is offered alongside these as an optional backend, and its LLM provider bills models at provider rates with no markup, with two DeepSeek models free when run there. No tier includes model inference, which is billed to your own provider key, at cost through OpenHands, or through an existing agent subscription.
Two things a buyer should weigh before adopting the hosted product. All Hands AI states in its privacy policy that it may use content and feedback from your use of the Services to train and tune its AI models. No mechanism to opt out of that is published. Its Enterprise Security FAQ states that for customers with an Enterprise Customer Agreement, customer data is not used to train or test its AI systems, and self-hosting the MIT build avoids the question entirely because nothing leaves your environment. Separately, the company runs a Vanta trust center that now lists SOC 2 under compliance and offers a SOC 2 Type II report on request, while its Security FAQ, still dated July 2026, records ISO 27001 as not yet held.
OpenHands suits teams that want to delegate whole engineering tasks to an agent they can inspect, self-host and run on their own model keys, and that have code review in place for the pull requests it opens. Devin at $20/mo and GitHub Copilot at $10/mo are commercial alternatives with managed infrastructure and support commitments, for teams that would rather not run model API spend or Docker infrastructure themselves. Teams that want a hosted agent that re-checks each task in a separate verifier session and runs under a hard dollar cap set in advance should look at Hyper.
Current state Q4 2026: the product ships as Agent Canvas, the Software Agent SDK and Agent Server, an Automation Server and a maintained CLI, with the old Local GUI deprecated. Agent Canvas is free, Teams is $500 a month and Enterprise is quote-only, and the trust center lists a SOC 2 Type II report, with a third-party penetration test completed in June 2026. A September 2026 Enterprise release added organization-wide secrets, a budget dashboard showing each member's usage, sign-in to MCP servers with the user's own identity through OAuth, including a catalog with GitLab and Atlassian Rovo, organization-wide visibility of automations, and Git Sync for managing automations through a Git repository. In September, Agent Canvas releases v1.17.0 to v1.24.0 added Agent Profiles that scope the tools, MCP servers and secrets each agent can use, an isolated Docker runtime for every conversation, and automation templates that chain issue triage, issue-to-pull-request and pull request review. In October 2026, v1.27.0 added DigitalOcean-hosted agents managed from Agent Canvas, and five OpenHands Enterprise releases through early November added a Super Admin dashboard and identity providers managed in the UI without Keycloak, both in beta, plus a guided first-time setup.