OpenHands

OpenHands is an open-source project for AI-assisted software development. The repository describes it as "AI-Driven Development", and it has more than 90,000 stars on GitHub. At its core is a software agent that does not merely suggest code but carries out tasks itself in an isolated environment: editing files, running commands in a terminal, executing tests, and checking the result. Anyone looking for an AI coding assistant that is tied neither to one model provider nor to one particular editor will find in OpenHands an approach that can be run entirely on your own infrastructure.

How OpenHands is put together

According to the documentation, OpenHands consists of several building blocks that can be combined depending on the use case. The Agent Canvas is an open-source browser client and a kind of control center for agent conversations; it connects to local or remote backends and, besides its own agent, can drive other coding agents that speak the Agent Client Protocol. The Software Agent SDK is a Python library for building your own agents, accompanied by an Agent Server with REST and WebSocket interfaces. The command line offers a terminal interface as well as a headless mode for scripts, automation, and CI/CD pipelines. This breadth is unusual: you can start small and later wire the same agent into a pipeline or a team system.

Sandbox and safety

An agent that runs commands needs boundaries. According to the documentation, OpenHands offers sandbox providers for this: a Docker sandbox, which is recommended for local use, a simple process mode, and remote environments. The README explicitly warns that without a sandbox the agent has full access to the file system. For sensitive projects you should therefore choose the Docker route. In addition, the SDK describes a confirmation policy with a security analyzer that controls which actions the agent may only perform after approval.

Model choice and MCP

According to the documentation, OpenHands can connect to any model reachable through LiteLLM. That includes OpenAI, Google Gemini and Vertex, AWS Bedrock, and local servers such as LM Studio, Ollama, or vLLM. Your own keys can be used in every plan. The agent can be extended through the Model Context Protocol: MCP servers can be added in the settings, on the command line, and in the SDK. Many entries from the Skill Road catalog are therefore also suitable for OpenHands, such as servers for databases, browsers, or developer tools.

Integrations for everyday teamwork

For the cloud variant, the documentation names integrations with GitHub, GitLab, Slack, and Jira Cloud. Tasks from a ticket system or chat can be handed to the agent, and the results collected again as changes in the repository. The Agent Canvas additionally mentions automations with Linear and Notion.

Plans and cost

The open-source entry point runs locally and is free; according to the pricing page it is MIT licensed. There is also a free Individual plan as a hosted service for one person with a limited number of daily conversations, cloud access, an API, and Jira and Slack integration. The Enterprise plan targets organisations and is available as a cloud service or self-hosted, with options such as a private VPC, SAML/SSO, and a named engineer; terms are individual. For model access you can bring your own keys or use the OpenHands provider, which according to the vendor bills tokens at cost with no markup. Current terms are on the official pricing page.

Where it sits

The difference from editor assistants such as Cursor or Zed is the shape: OpenHands works more as a standalone agent with its own environment that you hand tasks to. Compared with goose, which mainly runs locally as a general-purpose agent, OpenHands emphasises software development, the SDK, and team integration. Compared with Google's Jules, it is open source and can be self-hosted.

Who OpenHands fits

OpenHands suits developers and teams who want to operate a software agent with their own model choice, a sandbox, and open source code, and who are willing to set up Docker and permissions themselves. If you want a purely managed solution without running anything yourself, you can try the cloud plan. A sensible way to start is the Docker variant with a small test repository, a connected model, and a clearly scoped task such as writing tests for a single function.