Claude Code vs OpenHands: What Proves a Change Is Ready?
Compare Claude Code and OpenHands on completion evidence, independent review and the path from agent implementation to merge.
Founder, Task Machine
Should Claude Code or OpenHands decide when a coding job has enough evidence to move toward merge? Claude Code's /goal evaluator judges a completion condition using conversation evidence. OpenHands documents a software factory that requires passing independent statuses for the exact revision before conditional merge.
Both products publish ways to check agent work, so the decision cannot rest on a claim that only one performs verification. A developer choosing between them should identify what gets checked, who can change the implementation and what happens when the commit changes after review. A test result for an earlier revision leaves a different question from a test result for the revision you intend to ship.
Claude Code separates goals from deeper review
Claude Code's /goal documentation describes a separate model checking a measurable completion condition after every turn. It can tell the coding session to continue or stop the session's goal. The evaluator judges conversation evidence and does not independently read files or run commands.
That boundary does not describe every Claude Code review feature. Its Code Review documentation covers hosted review on Team and Enterprise, with repository guidance supplied through CLAUDE.md and review-specific REVIEW.md rules. Code Review chooses its own Claude models and ranks individual findings by severity.
For another review path, ultrareview launches a separate reviewer fleet in a cloud sandbox and independently reproduces and verifies findings. It is research preview. If independent review is your deciding requirement, evaluate this published capability without confusing it with the narrower /goal completion check.
OpenHands documents a revision-bound factory
OpenHands' software factory guide describes triage, development, independent review and conditional merge as four automations. The reviewer uses a separate token that cannot push code. It reviews and tests the exact pull-request revision, separating the review role from implementation rights.
A changed commit needs fresh acceptance. Failed work returns for revision, and the watchdog requires current independent acceptance before merge. That provides a concrete model to inspect if your concern is an agent amending a branch after a reviewer approved an earlier version.
OpenHands also offers a simpler goal loop in Agent Canvas conversations. An agent and judge work toward an objective until completion or an iteration limit, with missing work sent back as feedback. The factory's exact-revision conditions and a conversation's completion score answer different operational questions within the same product.
Compare orchestration artifacts, not saved prompts
Claude Code's dynamic workflows express loops, branches and agent stages in JavaScript. You can inspect and rerun the script, and independent agents can adversarially review one another's findings. A developer comfortable owning orchestration code can examine which stage produces evidence and which stage consumes it.
OpenHands' automation guide describes scheduled runs and matching webhook or service events. Script automations can execute a saved bundle directly without starting an agent conversation. The Automate view retains configuration, recent run history and status.
For an unattended repair job, both products publish more than a manual chat interface. Claude Code also has research-preview scheduled cloud tasks. Those runs open separate sessions, but the scheduled task documentation says a Succeeded status means the session ended without a reported failure, not that it completed the requested work.
OpenHands has a different history boundary to check. Its automation documentation says Cloud script logs become unavailable after the sandbox is deleted and advises opening them while the run is active. If a reviewer will inspect evidence later, include log availability in the evaluation of the exact automation type you plan to run.
Execution location and payment remain separate choices
Claude Code runs in local terminals and editors as well as cloud sessions. Public-beta Projects can move a thread to your computer through Remote Control for local files or tools. The Projects guide also says the feature is rolling out gradually on Pro, Max and Ultra, so availability needs checking before you depend on that handoff.
OpenHands' Canvas overview describes backends on your computer, in Docker, on a virtual machine or in managed Cloud. Canvas accepts provider API keys and can reuse a signed-in ACP agent subscription when the backend shares the agent's machine. Local access is therefore not the deciding difference between these products.
Claude's pricing page lists Pro at $20 billed monthly or $200 billed annually, Max at $100 per month and Ultra at $200 per month, excluding applicable tax. Claude Code also supports token-based API billing. Hosted Code Review uses organization usage credits separately from a plan's included usage, without a flat per-review price.
OpenHands' pricing page lists Canvas as free to self-host. Self-hosted Teams costs $500/month for up to 10 users, plus $50/month per additional seat, with a 5,000-session monthly cap. Cloud model usage uses credits, and the public pricing page does not give a managed Cloud plan amount. Separate identities and role-based access belong to Teams, while free Canvas uses a single user or shared identity.
Claude Code and OpenHands, point by point
The table compares answers to the same questions, each with its source and check date. Read the completion and review entries alongside the execution environment you intend to use.
Both columns answer the same questions from each company's public pages. Yes means the company publishes the capability the question asks about, Partly means something narrower, No means it says it does not offer it, and Not published means its public pages do not describe it. Every answer links to its source and shows when we last checked it. The oldest check on this page was November 13, 2026.
| What you compare | Claude Code | OpenHands |
|---|---|---|
| Goals |
Claude Code
Yes
The /goal command gives Claude a measurable completion condition, such as passing tests or an empty issue backlog. A separate model checks the condition after each turn and tells Claude to continue or stops the session's goal. |
OpenHands
Partly
Canvas runs an objective through an agent and a judge until completion or an iteration limit. Its docs describe completion scores rather than an owned goal with a tracked metric or target. |
| Leads |
Claude Code
Yes
Projects give Claude a coordinator role that starts and tracks threads, while experimental agent teams have a lead that assigns tasks and synthesizes results. Projects are public beta for Pro, Max and Ultra, with a gradual rollout. |
OpenHands
Partly
The daily-work agent prioritizes development work and dispatches separate agents, while factory triage checks issue dependencies. Its docs describe these workflows rather than named lead roles over projects, goals or teams. |
| Tasks |
Claude Code
Partly
Experimental agent teams share tasks with agent owners, dependencies and completion states, and Projects retain work in threads. Its docs do not describe the same work item being owned interchangeably by a person or an agent. |
OpenHands
Partly
Conversations retain their own history and agent configuration, and the daily workflow gathers assigned GitHub issues. Its docs do not establish a shared work record that either a person or an agent can own. |
| Recurring work |
Claude Code
Yes
Research-preview scheduled tasks run from recurring schedules, HTTP API calls or GitHub events. Each run opens a separate cloud session, with its history available from the saved task. |
OpenHands
Yes
Canvas automations run on schedules or matching webhook and service events. The Automate view shows their configuration, recent run history and status. |
| Workflows |
Claude Code
Yes
Dynamic workflows put loops, branches and agent stages in a JavaScript script that you can read and rerun. The script controls what runs next rather than leaving each orchestration decision to the conversation. |
OpenHands
Yes
The software factory defines triage, development, independent review and conditional merge as four automations. Script automations can execute a saved bundle directly instead of starting an agent conversation. |
| Autonomy |
Claude Code
Yes
Local permission rules allow, ask or deny tool actions, with approval prompts in the session. Public-beta Projects list threads needing approval under Waiting on you, although cloud threads do not apply repository permission rules. |
OpenHands
Yes
SDK confirmation policies require approval for all actions, risky actions or none. Pending actions are collected for confirmation and can be rejected with feedback through a custom confirmation handler. |
| Learned autonomy |
Claude Code
Partly
The /auto-mode-setup command drafts environment and rule entries from project configuration and commands in recent sessions, for you to accept. Its docs do not describe learning permission-change proposals from your approval history. |
OpenHands
Not published
Its public docs describe configured confirmation policies and feedback-led skill improvement. They do not describe proposing permission changes from your approval history. |
| Memory |
Claude Code
Yes
Auto memory keeps corrections and preferences across local sessions in files you can read, edit or delete. Public-beta Projects also have editable project memory that later cloud threads read. |
OpenHands
Yes
Opt-in persistent memory stores agent-maintained user and project notes in Markdown files loaded into later conversations. You can review, edit or delete those files. |
| Planning |
Claude Code
Yes
Plan mode researches a request and writes a proposed plan before implementation. You can revise the plan in your editor or ask Claude to keep planning before approving execution. |
OpenHands
Yes
The daily-work guide has you discuss a failing check, ask for a proposed fix and approve its plan before implementation. That review boundary is expressed through conversation instructions. |
| Intake |
Claude Code
Yes
An HTTP POST can send an incoming alert to a saved Claude Code task whose prompt tells the agent how to investigate it. GitHub events can also start these research-preview cloud runs. |
OpenHands
Yes
Signed webhook payloads are routed to automations by event name and optional filters, then added to the agent's prompt. Automations can select an Agent Profile on supporting backends. |
| Connectors |
Claude Code
Yes
MCP servers connect external tools and APIs. Research-preview scheduled tasks accept HTTP API triggers and GitHub events, and let you select which connectors each run can use. |
OpenHands
Yes
Canvas connects MCP servers for external tools and data, while custom webhook sources accept HTTP payloads and trigger matching automations. Some service signature formats need a validating relay. |
| Models |
Claude Code
Yes
The opusplan setting automatically uses Opus for planning and Sonnet for execution. Code Review also chooses its own Claude models rather than asking administrators to select one. |
OpenHands
Yes
The SDK's optional task-routing tool uses a classifier model to select a saved model profile from the current conversation. The agent decides when to request a switch, and routing configuration defines the candidates. |
| Cloud or your own computer |
Claude Code
Yes
Claude Code runs locally in your terminal and editors or in cloud sessions. Projects can send a thread to your computer through Remote Control when it needs local tools or files. |
OpenHands
Yes
Canvas can connect to a backend on your machine, in Docker, on a virtual machine or in managed OpenHands Cloud. You choose the execution environment and the workspace it can access. |
| Your own subscription |
Claude Code
Yes
You can sign in with a Claude subscription, supply an Anthropic API key or configure supported cloud-provider credentials. The published connections cover Claude usage rather than ChatGPT or GitHub Copilot subscription allowances. |
OpenHands
Yes
Canvas accepts provider API keys and can reuse a signed-in ACP agent subscription when the backend shares its machine. The SDK also documents ChatGPT Plus or Pro access to Codex models. |
| Pricing |
Claude Code
Yes
Claude Code supports token-based API billing and subscriptions with usage limits. Pro costs $20 billed monthly or $200 billed annually, Max $100 per month and Ultra $200 per month, excluding applicable tax. |
OpenHands
Yes
Cloud model usage uses credits. Canvas is free to self-host. Self-hosted Teams costs $500/month for up to 10 users, plus $50/month per additional seat, with a 5,000-session monthly cap. Enterprise has custom pricing. |
| Correctness |
Claude Code
Yes
The /goal evaluator checks a stated completion condition after each turn, such as tests passing, and sends unmet work back. It judges the conversation evidence and does not independently run commands or read files. |
OpenHands
Yes
Goal runs use a judge to check completion and send missing work back to the agent. The software factory establishes acceptance criteria, runs relevant tests and requires passing independent statuses for the exact revision before merging. |
For software work
| What you compare | Claude Code | OpenHands |
|---|---|---|
| Code review rules |
Claude Code
Yes
Code Review flags changes that violate CLAUDE.md instructions and accepts review-specific rules in REVIEW.md. The hosted service is available on Team and Enterprise plans. |
OpenHands
Yes
You can add repository-specific review rules as a skill that runs alongside the default code-review framework. The documented example covers approval decisions, lint conventions and test locations. |
| Adversarial review |
Claude Code
Yes
Research-preview ultrareview launches a separate fleet of reviewers in a cloud sandbox, with findings independently reproduced and verified. Dynamic workflows can also have independent agents adversarially review each other's findings. |
OpenHands
Yes
The software factory uses an independent reviewer with a separate token that cannot push code. It reviews and tests the exact pull-request revision, and the watchdog requires current independent acceptance before merge. |
| Risk rating |
Claude Code
Not published
Its docs describe no overall risk rating for a revision before merge. Code Review ranks individual findings by severity and leaves the merge decision with your team. |
OpenHands
Partly
The SDK rates individual actions LOW, MEDIUM, HIGH or UNKNOWN before execution. Its docs describe action security ratings rather than a risk rating of the whole change before merge. |
| Pull requests |
Claude Code
Yes
Public-beta Projects open pull requests from cloud threads and watch them with auto-fix. Threads push fixes for failed checks and address review comments before reporting the pull request ready for you. |
OpenHands
Yes
Factory agents turn GitHub issues into pull requests, receive native reviews and revise failed changes. A changed commit needs fresh acceptance, and the watchdog can merge once its required checks pass. |
A fact we cannot find or cannot check is left out, and prices can change after the date shown. Task Machine publishes this comparison and is an interested party. Claude Code and OpenHands are named only to identify them, and there is no relationship between the companies. Has something changed? Tell us.
Who each one suits
Pick Claude Code if you want coding sessions with a goal evaluator, scripted orchestration and separate review options. Its honest limit is that /goal evaluates conversation evidence, while ultrareview remains research preview and should be evaluated as a distinct feature.
Pick OpenHands if a factory with an independently reviewed exact revision matches your delivery process. Check its operating boundaries, including separate-identity pricing and the lifetime of Cloud script logs, before treating every automation's history as durable review evidence.
Where Task Machine fits
Task Machine defines acceptance requirements on each Task before execution and checks delivered work against them before human review. Its Task review documentation also describes adversarial review for an approved plan with high sensitivity or a Critical review level. Task Machine is the wrong pick if your need is already met by a repository-centered OpenHands factory or Claude Code's direct coding and review surfaces.
The narrower question is whether that coding work also needs a shared Task owned by a person or an agent. Inspect Task Machine vs Claude Code and Task Machine vs OpenHands, then choose one pending change and document which evidence must be refreshed after its next commit.