All Hands AI

What it is

All Hands AI is the company behind OpenHands, which it describes as the open platform for agentic software development^[1]. The framing matters: OpenHands is presented not as a single chat assistant but as a place where agents work on software. On the project's site, agents operate from conversations that are tied to repositories, and the recommended setups include a Slack channel monitor and a pull request review^[1]. In other words, the unit of work is a task in a codebase, and the conversation is the interface to it.

Where it came from

OpenHands has a public research record. It was published as arXiv preprint 2407.16741 on 23 July 2024 and accepted at ICLR 2025^[3]. The paper carries 24 authors, among them Xingyao Wang and Graham Neubig^[3]. The same record documents the project's open-source footing: at publication it was released under the MIT license, with more than 2.1K contributions from over 188 contributors^[3]. Those numbers describe a project that was already a community effort rather than a single team's internal tool when the paper appeared.

How it is put together

The documentation splits the ecosystem across separate repositories rather than shipping one monolith. It lays out Agent Canvas, a Software Agent SDK with an Agent Server, an Automation Server and a Sandbox Server, each in its own repository^[2]. The division follows the parts of the job: an interface where a person works with an agent, a toolkit for building agents, a server that runs them, a server that fires them off on triggers, and a sandbox in which the code they write can be executed in isolation. The documentation also notes that each public repository carries its own license rather than one licence covering the whole ecosystem^[2], so anyone reusing the pieces needs to check the terms repository by repository.

There are two ways to run it. One is the managed OpenHands Cloud; the other is an enterprise edition for licensed self-hosting or managed deployment^[2].

What people use it for

Automations can be triggered from GitHub, Slack, Jira, Linear and schedules^[1], which puts the agent in the places where software teams already raise and track work. The site's templates cover pull request review, Linear issue triage, Jira incident triage, security remediation and CI failure fixes^[1]. Read together, those templates describe the chores that recur in a repository: someone opens a pull request and it needs a first pass; an issue arrives and needs to be sorted; an incident is filed and needs triage; a dependency needs patching; a build breaks and needs a fix. These are bounded, repeatable tasks with a checkable outcome, which is a different proposition from asking a model to write a feature from scratch.

Where it sits

OpenHands belongs to a shift in how coding agents are packaged. Instead of a tool that suggests text inside an editor, it is an agent that acts on a repository: it reads, edits, runs commands and reports back. That shift brings its own engineering problems — where the agent's code runs, how a run is triggered, how a team builds its own agent on top — and the repository split described in the documentation is an answer to them^[2]. The MIT license and the contributor count at publication^[3] also place it in the open-source tradition rather than in a closed product line, even as the company offers a managed cloud and an enterprise edition alongside it^[2].

What the platform does not settle by itself is scope. An agent that can be pointed at a repository and given triggers from GitHub, Slack, Jira, Linear or a schedule^[1] is only as useful as the tasks it is given and the review that follows. The templates on the site are best read as a statement of where the builders think that works today: triage, review, remediation and build repair^[1].