OpenHands Review 2026: It Stopped Being Just an Agent and Became the Control Room
OpenHands launched in 2024 as "OpenDevin" — an open-source answer to Devin. By 2026, its own GitHub README doesn't call it an autonomous agent first anymore. It calls the flagship project "Agent Canvas": a self-hosted control center that can run OpenHands' own agent, or Claude Code, or Codex, or Gemini, side by side. This review covers what OpenHands actually is now, what it costs, and whether that pivot makes it more or less useful to you.
| Still free/open-source? | Yes — MIT license |
| Flagship repo name (2026) | Agent Canvas |
| Can it run other agents? | Yes — Claude Code, Codex, Gemini |
| Free tier limit | 10 conversations/day, 1 user |
| Funding | $23.8M, led by Madrona (2025) |
OpenHands didn't just survive its own hype cycle. It repositioned around what actually worked.
"OpenHands started as a bet that an open-source community could out-build Devin. In 2026 the company's real bet looks different: instead of trying to out-autonomy Devin, Claude Code, and Codex, OpenHands built the control room all of them can run inside — including its own agent, and including its rivals'."
If your last mental model of OpenHands is "OpenDevin, the open-source Devin clone," the honest update is: the underlying agent still exists, still writes code and opens PRs, and is still free to self-host. What's new is the framing — the flagship product is now a platform (Agent Canvas) that treats its own agent as one option among several, not the whole pitch.
One control plane, any agent underneath.
Agent Canvas is the self-hosted developer control center: a GUI, CLI, and SDK that can run the OpenHands agent out of the box, or point at Claude Code, Codex, Gemini, or any Agent-Client Protocol (ACP)-compatible agent instead. It runs locally by default, but can connect to remote or cloud "agent backends" — Docker containers, VMs, company infrastructure, or OpenHands' own Cloud/Enterprise hosting.
The practical read: OpenHands (the agent) did not disappear. It became the default option inside a bigger product (Agent Canvas) whose real differentiator is running whichever agent you already trust.
The star count is real; what it's measuring has shifted under it.
OpenHands/OpenHands, verified on GitHub, Sep 20, 2026.
Verified on GitHub, Sep 20, 2026.
$18.8M Series A (Madrona, Nov 2025) + earlier pre-seed.
More permissive than Cline's Apache-2.0.
| Era | What OpenHands was | What changed it |
|---|---|---|
| 2024 | "OpenDevin" — open-source community project answering Devin | Renamed OpenHands; company All Hands AI founded around it |
| 2025 | Autonomous "AI software engineer": issue-to-PR, self-hosted or hosted | $18.8M Series A (Madrona-led); Growth pricing tier existed |
| 2026 | Agent Canvas: control plane running OpenHands or rival agents | Company renamed itself OpenHands; Growth plan discontinued; GitHub org moved to OpenHands/OpenHands |
The praise predates the pivot; the caution is about scope, not the pivot itself.
"I haven't tried Devin yet but I freakin' love OpenHands... Just opening GitHub issues and the AI figures it out and writes tests and then pushes a PR is magical."
"OpenHands is a capable autonomous worker for scoped, verifiable tasks and an unreliable one for open-ended design." "Fix the login bug" wanders; "Make test_login_rejects_expired_token pass" converges.
88,614 stars and 11,654 forks with contributions credited from engineers at AMD, Apple, Google, Amazon, Netflix, and NVIDIA — a proxy for serious developer engagement, not a corporate endorsement.
The free tier is real but thin; the middle tier just disappeared.
Open Source
Run locally or on your own infra.
- Full Agent Canvas stack
- Community support
- Model-agnostic
Individual (SaaS)
1 user, up to 10 conversations/day.
- Hosted GUI, mobile access
- Jira/Slack integrations
- BYOK or at-cost model access
Enterprise
SaaS or self-hosted VPC.
- SAML/SSO, unlimited conversations
- RBAC, named Customer Engineer
Source: openhands.dev/pricing, verified Sep 20, 2026.
The right question isn't "agent or platform" — it's "how well-scoped is your task?"
This is where OpenHands' own agent is reported to work best — try the free Open Source or Individual tier.
That's specifically what Agent Canvas is built for in 2026 — its actual differentiator now.
There's no middle plan anymore — you're evaluating Enterprise, custom-quoted.
Enterprise supports self-hosted VPC with a published Vanta Trust Center.
Expect more supervision than the "autonomous" framing implies — see Limitations.
The thing OpenHands is still best known for: closing the loop from GitHub issue to draft PR.
Point OpenHands at a GitHub issue and it can write code, run terminal commands, browse documentation, call APIs, and open a pull request — the same "magical" loop the real Genco quote above describes. The Agent Canvas frontend adds a visible session view so you can watch the agent work rather than only see the final diff.
- Scoped, verifiable tasks with a clear pass/fail signal (a failing test, a known CVE).
- Reviewing pull requests, investigating CI failures, updating dependencies.
- Reading and browsing documentation/APIs as part of completing a task.
- Open-ended, ambiguous asks ("fix the login bug") without a way to self-check.
- Full sandboxed Docker execution means the agent has real filesystem/command access if you skip the sandbox warning in the docs.
Issue in, verifiable change out — ideally with a test that proves it.
Official OpenHands video introducing the open-source AI software development agent.
Published by the official "OpenHands" YouTube channel (youtube.com/@OpenHands-AI). A separate third-party video ("OpenHands: Self-Host Your Own AI Dev Team," by channel "Elestio") was found during research and deliberately excluded as non-official.
The same engine, three doors: GUI, CLI, and SDK.
Beyond the Agent Canvas GUI, OpenHands ships a feature-complete CLI for terminal workflows and a composable Python Agent SDK for embedding agentic development into other tools. Enterprise/Cloud tiers add prebuilt automations that trigger on a schedule or webhook — generating a report and posting it to Slack, or auto-decomposing a GitHub issue into sub-tasks, without a human kicking off each run.
This is the actual 2026 story: OpenHands runs its rivals.
Agent Canvas supports the Agent-Client Protocol (ACP), meaning it can launch and manage sessions for OpenHands' own agent, Claude Code, Codex, Gemini, or any other ACP-compatible agent, switching between local, remote, and cloud backends without losing your place. This is a genuinely unusual competitive stance: most vendors want you locked to their own agent.
Your model bill is separate from your OpenHands bill, same as most BYOK agents.
| What you're doing | What you pay |
|---|---|
| Self-hosted Open Source | $0 to OpenHands; your model provider's own rate (BYOK) |
| Individual/SaaS (hosted, free) | $0 to OpenHands for the plan; BYOK, or pay-as-you-go via the OpenHands provider at cost, "no markup" |
| Enterprise | Custom quote; includes model-agnostic support for Anthropic, OpenAI, Bedrock, or others |
The self-hosting story is genuinely strong; the certification story is a Trust Center, not a named badge.
Teams with well-defined, checkable work — or a multi-agent stack to unify.
| Situation | Why OpenHands fits | Likely plan |
|---|---|---|
| Solo developer with scoped, testable tasks | Issue-to-PR loop works best when success is checkable | Open Source or Individual (free) |
| Team already running Claude Code and/or Codex | Agent Canvas orchestrates all of them from one place | Individual or Enterprise, depending on scale |
| Regulated org needing private VPC deployment | Self-hosted Kubernetes, no shared tenancy, Vanta Trust Center | Enterprise, custom |
| Team needing >10 conversations/day or >1 user today | No self-serve middle tier exists since Growth was discontinued | Enterprise (sales conversation required) |
Four real gaps, not manufactured ones.
Most secondary sources online (reviews, wikis, comparison sites) still describe OpenHands purely as an autonomous agent; the Agent Canvas pivot is easy to miss unless you read the live README.
10 conversations/day and 1 user is enough to evaluate, not enough to run daily team workflows.
The Growth plan's 2026 discontinuation leaves a real gap between free and custom-quoted Enterprise.
OpenHands' homepage lists Google, NVIDIA, Apple, MongoDB, Red Hat, Netflix, C3.ai, Amazon, Roche, VMware, and TikTok. These are vendor-published claims; JAVIS did not independently verify deployment scale at any of them.
If OpenHands' scope doesn't match your task, here's how to think about it.
| If you mostly need... | Compare OpenHands with... |
|---|---|
| Editor-first, approval-at-every-step control (not full autonomy) | Cline — read our review |
| The most widely deployed IDE-native assistant with native GitHub PR flow | GitHub Copilot — read our review |
| The closed, commercial autonomous agent OpenHands was built to rival | Devin (devin.ai) |
| A terminal-first agent OpenHands can itself orchestrate via ACP | Claude Code — see our Claude Code setup guide |
No affiliate relationship shapes this review.
JAVIS has not established a publisher affiliate program with OpenHands. Every CTA here points to the official product.
Go to the official page.
Open OpenHandsRead the alternative that matches your real question.
Choose the next articleOpenHands makes most sense next to the agents it can now run itself.
Questions people are actually searching right now
Is OpenHands the same as OpenDevin?
Yes — OpenDevin was the original name (launched Mar 2024) before renaming to OpenHands on Aug 26, 2024. The underlying lineage is continuous.
Is OpenHands still free?
Yes. The Open Source tier is free to self-host, and the hosted Individual/SaaS tier is also free, capped at 1 user and roughly 10 conversations/day.
What is Agent Canvas?
OpenHands' flagship 2026 product: a self-hosted developer control center that can run the OpenHands agent, or third-party agents like Claude Code, Codex, and Gemini, via the Agent-Client Protocol (ACP).
Did the company "All Hands AI" disappear?
No — it renamed itself to OpenHands, matching the product name, consistent with its own org profile.
What happened to the Growth pricing plan?
It was discontinued in 2026, per pricing trackers cross-checked against the current official pricing page, leaving Individual (free) and Enterprise (custom) as the two live tiers.
Can OpenHands run Claude Code or Codex?
Yes — this is the core of the Agent Canvas pivot: it's built to run any ACP-compatible agent, not only its own.
Is OpenHands better than Devin?
Different trade-offs: Devin is closed and cloud-run; OpenHands is open-source, self-hostable, and (per independent review) more reliable on scoped, verifiable tasks than open-ended ones.
Is OpenHands secure enough for enterprise use?
It supports self-hosted VPC/Kubernetes deployment with no shared tenancy and publishes a Vanta-backed Trust Center; JAVIS did not find a named SOC 2/ISO certification badge on official pages as of this check — verify current status directly before a regulated rollout.
Where this came from, and when it was checked.
GitHub stats (stars/forks/license): GitHub data for github.com/OpenHands/OpenHands, Sep 20, 2026.
Agent Canvas pivot / ACP multi-agent support: raw.githubusercontent.com/OpenHands/OpenHands/main/README.md, official, read directly.
Pricing (Open Source, Individual, Enterprise, discontinued Growth): openhands.dev/pricing, official, cross-checked against independent pricing trackers.
Funding ($23.8M, Madrona Series A): startupintros.com funding profile, corroborated by original TechCrunch pre-seed coverage.
Security/compliance posture: openhands.dev/enterprise, official, including the live Vanta Trust Center link.
Real screenshots/media: openhands.dev social-preview image, assets.openhands.dev screenshot and logo, opengraph.githubassets.com repo card.
Official video: youtube.com/watch?v=oB4JR98KRAA, authorship confirmed on YouTube (author "OpenHands", channel @OpenHands-AI); a non-official third-party video was found and deliberately excluded.
Real user quote: x.com/cgenco/status/1887551887001424102, public, attributed, linked directly.
Independent review pattern: theaiagentindex.com/agents/openhands.
