OpenHands self-hosted AI agent: what it does and where it fits
OpenHands is an open-source AI agent platform that a business can run on its own infrastructure. Here is how it works and how MILabs sets it up.
Updated 2026-09-21
What OpenHands is
OpenHands is an open-source platform for AI agents that carry out software tasks. All Hands AI develops it, and it is typically run through Docker on a server or computer the user controls. A person describes a task in plain language, and the agent works through it inside an isolated environment, editing files, running commands and reporting back on each step. A web interface and a command line let people follow progress and step in when needed. Because the software is self-hosted, the business decides where it runs and which language model it connects to. A hosted chat service, by contrast, does its work on the vendor's systems. The platform is aimed mainly at technical teams, but the same approach can support internal tasks that involve files, scripts or repeatable steps. Its value depends on clear instructions, a defined scope, and a person who reviews the output before it reaches customers or live systems.
OpenHands in plain language
A chat assistant replies to a question. An agent is given a goal and decides each next step from what it finds, such as opening a file, reading an error and trying a fix. The language model supplies the reasoning, while the surrounding software supplies the ability to read folders, run commands and call tools.
Each step can change something, so the real question for an owner is where the agent is allowed to act. Sandboxing, a contained space where commands run apart from the rest of the machine, is a key safeguard, which is why the environment matters as much as the model.
Because the code is open, a team can read how the agent behaves, adjust it, and choose when to take updates. That freedom carries upkeep. Someone must pick the model provider, decide which credentials the agent holds and watch the logs.
It helps to separate three parts: the agent platform, the language model it calls, and the project it works on. Each can change independently, and deciding on each is the first step toward a sensible scope.
What businesses use it for
- Drafting scripts that move data between spreadsheets and internal systems.
- Writing small internal tools, such as a form that tidies incoming records.
- Making repetitive changes across a set of files, such as renaming or reformatting exports.
- Reviewing a codebase to answer a specific question, such as where a setting is defined.
- Testing a proposed change in an isolated environment before it reaches the live system.
- Producing a first draft of documentation for an existing internal process.
Who it suits
- Technical teams that want an agent for code and script tasks on infrastructure they control.
- Businesses with private code or data that should stay off hosted services.
- Owners who have an engineer or IT provider available to review each change before release.
How MILabs sets it up
MILabs starts with the owner or team lead to agree what the agent should do and what it must never touch. Together we choose how OpenHands is deployed, which machine hosts it and which language model provider it connects to. We then connect the systems the business already uses, such as its source code repository, file storage and any internal database, using access no broader than the task requires. Permissions are set so the agent can read and change only the projects it needs, and every proposed change is reviewed by a person before it is accepted. We also agree who looks after the installation, who applies updates, who holds the API keys and who receives alerts. The business keeps a short written note of these decisions.
Builds it fits
Industries it fits
OpenHands in practice
In a working business, the agent rarely runs on its own initiative. Work usually starts with a written task, such as changing how an export is formatted. The person who wrote it stays responsible, and the agent proposes a change rather than deciding it is correct.
A typical session runs like this. The agent reads the relevant files, plans its steps and edits them in the isolated workspace. The reviewer then checks the changed files and the commands run, applies the team's usual tests, and accepts the change or returns it with a note. Separating requester from reviewer makes quiet errors easier to spot.
Teams tend to keep tasks small. A narrow request, with a clear list of files it may touch, is easier to check than an open-ended instruction to improve a system. When the agent stalls, stop the run, read the log, and either narrow the task or finish the step by hand.
Over time, a business gathers a short list of approved task types, each with the permissions it needs and a named checker. That list is what makes the work repeatable.
Frequently asked questions
- What is OpenHands?
- OpenHands is an open-source platform from All Hands AI for running AI agents that carry out software tasks. A user describes the task, and the agent plans steps, edits files, runs commands and reports results inside an isolated environment. Because it is self-hosted, the business chooses where it runs and which language model it connects to. It fits technical work best, where a reviewer checks each change before it reaches a live system.
- Which builds work well with OpenHands?
- Files, scripts and internal data suit OpenHands best. The private document question-and-answer server and document filing automation fit when documents must stay on the business's own infrastructure. A staff knowledge assistant or policy compliance assistant can draw on internal material the agent reads and summarises. CRM data entry automation works where the agent moves records between a spreadsheet and a system through scripts. Each still needs a named reviewer before release.
- How does MILabs set up OpenHands?
- MILabs starts with the owner or team lead to agree the purpose of the agent and the limits on what it may access. We choose the deployment, connect the systems the business already uses, and set permissions so the agent reaches only the projects it needs. We also agree who maintains the installation, applies updates and receives alerts. Each change the agent proposes is reviewed by a person before it is accepted.
- Where does OpenHands store data?
- Storage depends on how the business deploys OpenHands. When it runs on a server or computer the business controls, project files and any logs or workspace files stay on that infrastructure. Requests to the language model provider are separate, because the prompts and any code included in them travel to whichever model API has been configured. MILabs records which provider is used and what it receives, so the owner knows where information leaves.
- When is OpenHands the wrong choice?
- OpenHands is a poor fit when a business wants a finished product with no technical oversight, or when nobody can review the agent's output. Work where every answer must be guaranteed suits it less well. Self-hosting adds maintenance, so a business without someone responsible for updates and security may be better served by a hosted service. In those cases, a workflow tool or hosted chat assistant is usually easier to run.
