OpenHands is an autonomous software-development platform with CLI, local web, headless, SDK, hosted, and self-hosted routes. Its core code and core agent-server images are MIT licensed, while its enterprise directory uses a separate enterprise license. The recommended local path uses a Docker sandbox, which limits the host boundary only to the extent that mounts, credentials, and network access are constrained. OpenHands is the better choice than Devin when self-hosting, model choice, inspectability, or platform customization is a requirement. Devin is the better choice when a managed environment, first-party ticket integrations, DeepWiki, Review, billing, and support should come from one vendor. Continuum belongs in a third category: a supervised workbench around existing native agents rather than an autonomous platform.
- OpenHands is a platform: CLI, local web, headless, SDK, hosted, and self-hosted paths around autonomous development work.
- The core and core agent-server images are MIT licensed; the enterprise directory has separate terms.
- Its recommended local route uses a Docker sandbox, but mounts, credentials, and network access still define the real blast radius.
- Free code is not free operation. Models, compute, storage, upgrades, secrets, monitoring, and support remain costs.
- Choose OpenHands over Devin for control and customization; choose Devin for managed integration and support.
- Continuum is relevant when the requirement is visible supervision of several native agents, not owning an autonomous-agent platform.
What OpenHands actually is
OpenHands is an autonomous development platform rather than a single terminal chat. It can run through a command-line surface, a local web interface, headless automation, an SDK, a hosted service, or infrastructure you operate. The common job is to give an agent a repository task and an execution environment, let it inspect, edit, run commands, and iterate, then return a result for review.
The project was previously known as OpenDevin, which explains why both names still appear in queries and old links. The current name is OpenHands. The conceptual positioning survived the rename: an open platform for the Devin-shaped loop, where the environment and agent runtime are explicit parts of the system rather than hidden behind one chat box.
The OpenHands surfaces documented in the site’s August 2026 coverage.
| Surface | What it is for | Who owns the runtime |
|---|---|---|
| CLI | Interactive or scripted agent work from a terminal | You, unless connected to hosted infrastructure |
| Local web interface | Browser-based operation on a local setup | You |
| Headless mode | Automation and issue-driven flows without an interactive UI | You or your platform |
| SDK | Embedding or composing the agent behavior in another system | Your application team |
| Hosted route | Use the platform without operating every component | OpenHands service |
| Self-hosted deployment | Control environment, network, models, storage, and policy | Your organization |
Calling it an open-source coding agent understates the scope. Aider and OpenCode are easier to understand as daily agent clients. OpenHands is closer to an application platform for autonomous jobs: environments, browser or network-capable tasks, headless execution, and programmable composition. That scope is why it is the plausible open-source Devin comparison and why it is excessive for someone who only wants help editing one function.
What open source covers, exactly
The core OpenHands code and the core agent-server images are MIT licensed in the site's existing source review. That permits inspection, modification, and redistribution of those components under a permissive license. The repository also contains an enterprise directory under a separate enterprise license. A public repository can contain more than one legal boundary, so “OpenHands is MIT” is directionally useful and insufficient for procurement.
| Claim | Supported interpretation | Unsupported shortcut |
|---|---|---|
| OpenHands core is open source | Core code can be inspected and modified under MIT | Every commercial component has identical terms |
| The platform is self-hostable | You can operate documented paths on your infrastructure | The deployment maintains and secures itself |
| There is no core seat fee | The software license does not create a per-user core charge | Models, hosts, storage, and staff cost nothing |
| The agent can use local models | Inference can be routed to infrastructure you control when configured | Every local model matches frontier reliability |
| Docker is the sandbox | The agent server can run within a container boundary | Every mount, secret, and network path is automatically safe |
An enterprise evaluator should inventory the exact images, directories, services, and SDK packages that will ship. Record the license for each, the source commit, the update owner, and whether any hosted dependency remains. The same exercise belongs in a security review because the component boundary and the trust boundary often align: the part that stores secrets or schedules jobs may not be the part described by the homepage badge.
Open source is valuable even when no fork is planned. It allows an organization to inspect permission handling, trace how commands reach the runtime, audit network integrations, pin a release, and repair a compatibility gap. Forkability is also an obligation. Once a private patch diverges, upstream security fixes and model changes become a merge problem for the team that wanted control.
The Docker sandbox is a boundary, not a force field
OpenHands recommends a Docker sandbox for local execution in the existing documentation coverage. That is a sound default because an autonomous development agent installs dependencies, runs arbitrary commands, writes files, may browse the web, and may talk to external services. A disposable container gives those actions a process and filesystem boundary separate from the host.
The boundary is only as narrow as its configuration. A repository mounted read-write can be changed or deleted. A home directory mounted for convenience exposes far more than the task requires. A token passed into the container can be used with every permission that token carries. Unrestricted outbound network access can move source or secrets outside the host. Docker reduces accidental reach; it does not override capabilities you deliberately hand through the wall.
| Control | Safer default | Failure if omitted |
|---|---|---|
| Filesystem mount | One disposable worktree read-write; references read-only | Agent can modify unrelated repositories or user files |
| Credentials | Task-scoped tokens with no merge or deploy authority | Prompt injection or a bad command can use broad account power |
| Network | Allow only destinations needed for the task | Source and secrets have an exfiltration path |
| Container lifetime | Destroy after the task and retain only reviewed artifacts | State drifts and later runs inherit unknown changes |
| Base image | Pinned, reviewed image with a scheduled update owner | Silent dependency and toolchain drift |
| Resource limits | CPU, memory, disk, runtime, and concurrency caps | A failed loop consumes the host or creates an unbounded bill |
| Git authority | Push to a task branch; no direct main or merge access | Agent can turn a coding mistake into a release event |
The repository should enter as a clean, isolated worktree or clone. That gives the agent a disposable branch and gives the reviewer a precise diff. Do not mount a developer's already-dirty checkout and expect the container to preserve provenance. The container can isolate processes while git history still mixes human and agent changes.
OpenHands versus Devin
OpenHands and Devin overlap on autonomous software work. Both can accept a task, operate in an environment, use tools, test changes, and return a reviewable artifact. The durable distinction is who owns the platform. Cognition owns and integrates the managed Devin service. An OpenHands adopter can choose the hosted route or assume control of models, environments, policy, deployment, and operations.
The honest comparison is platform ownership, not a transient model leaderboard.
| Question | OpenHands | Devin |
|---|---|---|
| Core product model | Open-source autonomous development platform | Managed autonomous software engineer service |
| Deployment | Local, self-hosted, or hosted routes | Cognition-managed environment plus local CLI paths |
| License visibility | MIT core; enterprise directory separate | Commercial product and service |
| Models | You configure supported inference routes | Cognition packages model access and quota |
| Sandbox | Docker is the recommended local path | Managed environment is part of the service |
| Ticket and chat workflow | Build or configure the integrations you need | Slack, Teams, Linear, Jira, and web workflow |
| Codebase knowledge | Platform and context configuration are yours | DeepWiki and Cognition product integration |
| PR review | Build the review path around returned changes | Devin Review and visual QA |
| Operations owner | You on self-hosted routes | Cognition for the managed service |
| Cost shape | Models, compute, storage, and operations | Subscription quota, credits, or enterprise ACUs |
Choose OpenHands when self-hosting is a requirement, the model route must remain under your control, the agent platform needs customization, or source-level inspection is part of policy. Choose Devin when the team wants one vendor to own the environment, account workflow, repository knowledge product, pull-request review surface, billing, and support.
Do not choose OpenHands merely to avoid Devin pricing. The current Devin Pro plan is $20 a month, while a self-hosted platform can consume many times that amount in engineering attention before the first reliable task. Conversely, do not choose Devin merely to avoid setup when code locality, custom policy, or provider independence is a hard requirement. Cheap managed service cannot satisfy a boundary it does not offer.
OpenHands versus a supervised agent workbench
OpenHands is also compared with products that look similar on a screenshot and solve a different problem. An autonomous platform owns the loop and environment for a delegated task. An agent workbench organizes sessions from other agents, keeps their branches and status visible, and lets a person intervene. Both may show terminals, diffs, tasks, and agents. The ownership boundary is the distinction.
| Question | OpenHands | Continuum workbench |
|---|---|---|
| What is the product? | An autonomous agent platform | A control plane around supported native coding agents |
| Which agent runs? | OpenHands runtime and configured models | Official Claude Code, Codex, Cursor, Grok, Gemini-side, and OpenCode paths documented by the site |
| Where does it run? | Docker, self-hosted, or hosted routes | Mac or enrolled Linux and Windows hosts; optional cloud is separate |
| Human loop | Policy and deployment decide intervention points | Plan, approve, interrupt, diff, PR, terminal, and artifact surfaces |
| Mobile control | Depends on the deployment and interface you build | Native iPhone and Watch plus web projection |
| Cost meter | Model, runtime, platform operations | Free app plus provider subscriptions or keys; hosted inference optional |
| Open source requirement | MIT core | Commercial product with a free app tier |
| Best reason to choose | Own and customize autonomous infrastructure | Supervise several existing agents across devices and accounts |
Continuum is relevant to this guide because teams searching for an open-source Devin sometimes want neither source code nor a managed autonomous service. They want to run several Claude Code or Codex sessions, keep each in a worktree, see what is waiting, approve a plan from a phone, and understand quota and cost. That is an orchestration requirement. Installing an autonomous platform to solve session visibility is choosing the wrong layer.
The real OpenHands cost
The repository can be free while the system is expensive. An autonomous task consumes inference, a runtime, storage, network, image pulls, logs, and review. A production self-hosted service also needs deployment automation, secret rotation, monitoring, upgrades, rollback, and an owner when provider APIs or model identifiers change.
| Cost line | Self-hosted OpenHands | Managed Devin comparison |
|---|---|---|
| Software license | $0 for MIT core | Included in commercial plan |
| Model inference | Provider API, gateway, or local hardware | Included quota plus on-demand credits |
| Runtime compute | Docker host, VM, cluster, or developer machine | Managed environment packaged by Cognition |
| Storage and logs | Your retention and backup design | Service behavior and contract |
| Security | Your mounts, tokens, network policy, image review, and incident response | Shared responsibility with the managed service |
| Updates | Your version pin, testing, rollout, and rollback | Vendor release and support path |
| Human review | Still required | Still required |
A local model can remove a per-token cash invoice after hardware exists. It can add latency, lower task reliability, and increase review or retries. A frontier API can finish faster while creating variable spend. The relevant measure is accepted work per dollar and per reviewer hour, not whether inference appears on a cloud bill.
Self-hosting is economical when it satisfies a hard constraint or reaches enough utilization to justify an internal platform. Code locality, model independence, custom policy, and integration with private infrastructure are coherent reasons. “The GitHub repository is free” is not. A small team with no operator should price the hosted OpenHands route and Devin alongside the self-hosted estimate rather than assuming the latter wins.
monthly cost =
model inference or local hardware allocation
+ runtime compute and storage
+ platform engineering and upgrades
+ security and incident-response allocation
+ human scoping, review, and rework
A safe first deployment
The first OpenHands task should be boring, reversible, and cheap to prove. Do not start by granting a persistent service write access to every repository. The goal is to validate the environment and review path before testing autonomy depth.
Pin the exact release and read its licenses
Record the OpenHands commit or image digest, identify which core and enterprise components are present, and name the person responsible for updates.
Create a disposable worktree
Mount only one clean task worktree read-write. Keep the main checkout, home directory, SSH material, and unrelated repositories outside the container.
Issue narrow credentials
Use a repository token that can read and push a task branch but cannot merge, administer, deploy, or reach other organizations.
Constrain the environment
Set CPU, memory, disk, runtime, and concurrency limits. Restrict outbound destinations to package registries and task services that are actually needed.
Choose a deterministic task
Use a reproducible bug or dependency update with exact tests and no external side effects. Name the conditions that require the agent to stop.
Review the trajectory and diff
Inspect unexpected commands, network use, dependency changes, deletions, and test evidence before accepting any result.
Destroy the runtime
Retain the branch, relevant logs, usage, and review record, then remove the task container and revoke its credentials.
Who should choose OpenHands
| Requirement | Best first choice |
|---|---|
| Self-hosted autonomous development is mandatory | OpenHands |
| Agent platform must be inspectable and modifiable | OpenHands core, after component-level license review |
| One vendor should operate environment, integrations, review, billing, and support | Devin |
| A developer wants a smaller open-source daily coding client | OpenCode or Aider |
| Several existing provider agents need visible worktrees, mobile control, and cost context | Continuum |
| The team has no operator and only a few autonomous tasks | Test hosted routes before self-hosting |
| The only reason is “free Devin” | Recalculate total cost before installing |
OpenHands is an unusually credible project when the organization wants to own the autonomous layer. It offers enough surfaces to be a platform rather than a demo, a documented container boundary for local execution, and permissive core code that can be inspected. Those are substantial advantages. They are not the same as a turnkey service.
Choose the smallest system that satisfies the hard requirement. If provider choice and a daily terminal are enough, OpenCode or Aider will be easier to operate. If asynchronous managed delegation is the requirement, Devin may be cheaper in total. If the requirement is a supervised fleet of agents already licensed elsewhere, use a workbench. OpenHands earns its complexity when the autonomous platform itself must belong to you.
Questions people ask
OpenHands is an autonomous software-development platform with CLI, local web, headless, SDK, hosted, and self-hosted routes. It can inspect repositories, edit code, run commands and tests, and return work for review.
OpenHands is the current name of the project previously known as OpenDevin. Older articles and queries still use the prior name.
The core code and core agent-server images are MIT licensed in the site’s current coverage. The enterprise directory carries a separate enterprise license, so check every component you plan to deploy.
The MIT core has no software seat fee, but model inference, Docker or VM compute, storage, operations, security, updates, and human review still have costs. Hosted terms are a separate question from the core license.
The recommended local sandbox path uses Docker. The real boundary depends on mounted files, injected credentials, network access, container privileges, and resource limits.
OpenHands is better when self-hosting, model choice, source inspection, or platform customization is required. Devin is better when one vendor should operate the managed environment, integrations, DeepWiki, Review, billing, and support.
OpenHands can be configured around local inference routes, but model compatibility and quality must be tested. Local inference removes a token invoice, not hardware, latency, maintenance, or review cost.
Only for a different job. Continuum supervises supported native agents with worktrees, plans, diffs, PRs, mobile control, quota gauges, and local spend. It is not an open-source autonomous runtime or a forkable agent platform.
Treating a container or an open-source badge as a complete security and operating model. Mounts, credentials, network access, updates, model spend, review, and incident response all need explicit owners.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.