← Back to research
·10 min read·opensource

OpenHands

OpenHands (formerly OpenDevin) is the most-starred open-source autonomous coding agent at 76K+ GitHub stars, spanning CLI, local GUI, cloud, and SDK surfaces — backed by an $18.8M Series A led by Madrona.

Key takeaways

  • 76K+ GitHub stars as of June 2026 — the most-starred open-source coding agent — backed by an $18.8M Series A led by Madrona (November 2025)
  • Full autonomy is the design point: agents write code, run commands, browse the web, and open PRs end-to-end in sandboxed Docker/Kubernetes environments, closer to Devin than to an IDE copilot
  • Model-agnostic (any LLM via BYOK or OpenHands' at-cost provider) with four surfaces — Python SDK, CLI, local GUI, and OpenHands Cloud — plus a source-available enterprise edition

FAQ

What is OpenHands?

OpenHands is an open-source platform for autonomous AI software agents that can write code, run commands, browse the web, and open pull requests — executing complete engineering tasks end-to-end rather than suggesting completions.

How much does OpenHands cost?

The open-source tool is free (MIT, except the enterprise/ directory). OpenHands Cloud's Individual plan is free to start with a 10-conversation daily cap; you bring your own LLM keys or pay OpenHands' provider at cost with no markup. Enterprise (SaaS or self-hosted) is custom-priced.

What models work with OpenHands?

Any LLM — the platform is model-agnostic via a LiteLLM-based provider layer, supporting Claude, GPT, Gemini, and local models with your own API keys; the free cloud tier runs on the Minimax model.

How is OpenHands different from Devin?

Both target autonomous end-to-end task execution, but OpenHands is open source, self-hostable, and model-agnostic, while Devin is a closed, subscription-priced managed service.

Executive Summary

OpenHands is an open-source platform for autonomous software agents — agents that write code, run commands, browse the web, and open pull requests end-to-end, rather than suggesting completions inside an editor. With 76,000+ GitHub stars and 9,700+ forks as of June 2026, it is the most-starred open-source coding agent, and the repository receives commits daily.[1] The project began in March 2024 as OpenDevin, an academic open-source answer to Devin co-created by CMU professor Graham Neubig, and grew into a company (All Hands AI, since rebranded to OpenHands) co-founded by Robert Brennan, Xingyao Wang, and Neubig.[2]

In November 2025 the company raised an $18.8M Series A led by Madrona, with participation from Menlo Ventures, Pillar VC, Obvious Ventures, Fujitsu Ventures, and Alumni Ventures, following an earlier $5M seed round.[3][2] The stated mission: "powerful coding agents for every developer, for free, forever" — funded by a source-available enterprise edition and a hosted cloud.[3]

AttributeValue
CompanyOpenHands (formerly All Hands AI)
Founded2024 (as the OpenDevin open-source project)[1]
Funding$18.8M Series A (November 2025, led by Madrona) + $5M seed[2]
GitHub Stars76,000+ (June 2026)[1]
LicenseMIT (except the source-available enterprise/ directory)[1]

Product Overview

OpenHands positions itself against code-suggestion tools: you hand the agent a task — a bug ticket, a vulnerability report, a migration — and it plans, edits files across the codebase, runs and tests the changes in a sandbox, and opens a reviewable PR.[4] Advertised production use cases include fixing vulnerabilities, reviewing pull requests, migrating legacy code (e.g., COBOL to Java), triaging incidents, and automating refactoring and test generation.[4]

Everything is built on the OpenHands Software Agent SDK, a composable Python library — define agents in code, run them locally, or scale to thousands of concurrent agents in the cloud.[1]

Key Capabilities

CapabilityDescription
End-to-End Task ExecutionPlan, write, run, and verify code changes autonomously
Command ExecutionRuns shell commands inside an isolated sandbox
Web BrowsingAgents browse the web to gather context
PR AutomationIssue-to-PR workflows via GitHub/GitLab integrations
Tool IntegrationsNative Slack, Jira, Linear, and CI/CD integrations (Cloud)[1]
Sandboxed RuntimeAgents run in isolated Docker/Kubernetes environments with auditability[4]

Product Surfaces

SurfaceDescriptionAvailability
CLITerminal agent, "familiar to anyone who has worked with Claude Code or Codex"[1]GA
Local GUISelf-hosted web app (REST API + React SPA), Devin/Jules-styleGA
OpenHands CloudHosted GUI with multi-user support, RBAC, and conversation sharing; free tier on the Minimax model[1]GA
SDKPython library powering all surfaces; embed agents in your own apps[5]GA
EnterpriseSelf-host OpenHands Cloud in your VPC via Kubernetes; source-available, license required beyond one month[1]Enterprise

Technical Architecture

OpenHands runs agents in isolated sandbox environments — Docker locally, Kubernetes at enterprise scale — so an autonomous agent executing arbitrary commands can't damage the host system.[4] The architecture is model-agnostic: any LLM works through a LiteLLM-based provider layer with your own API keys, or through OpenHands' provider billed at cost.[6]

The README advertises a 77.6 score on SWE-Bench, continuing the project's history as one of the strongest open-source performers on the benchmark since its OpenDevin days.[1]

Key Technical Details

AspectDetail
DeploymentLocal (CLI, Docker GUI), hosted cloud, or self-hosted VPC (Kubernetes)
Model(s)Any provider via BYOK/LiteLLM — Claude, GPT, Gemini, local models; free cloud tier uses Minimax[1]
IntegrationsGitHub, GitLab, Slack, Jira, Linear, CI/CD[4]
Open SourceYes — MIT, except the source-available enterprise/ directory[1]

Strengths

  • Category leader by adoption — 76K+ stars as of June 2026, the most-starred open coding agent, with daily commit activity[1]
  • True autonomy — Designed for delegation (ticket in, PR out), not supervised pair programming; one customer claims it "autonomously fixes 87% of bug tickets same-day"[4]
  • Every surface covered — SDK, CLI, local GUI, hosted cloud, and self-hosted enterprise from one codebase[1]
  • No vendor lock-in — Model-agnostic, MIT-licensed core, self-hostable, full code and data ownership[7]
  • Strong benchmark pedigree — Academic roots (CMU's Neubig) and a long SWE-bench track record; README badge shows 77.6[1]
  • Funded with a clear model — $18.8M Series A plus a source-available enterprise tier and at-cost cloud inference[3][6]

Cautions

  • Setup friction — Self-hosting demands Docker-in-Docker, API key configuration, and environment tuning; an April 2026 review calls it "not a simple download-and-run experience"[7]
  • Frontier-model dependent — Performance drops significantly with weaker LLMs, so realistic use requires expensive frontier models[7]
  • Agent loops — Reviewers report the agent sometimes repeats the same failing approach until a human intervenes with better instructions[7]
  • Weak on frontend work — UI code generation is unreliable because the agent struggles with visual requirements it cannot see[7]
  • API costs add up — Roughly $0.15–$0.60 per task in API fees when self-hosting; managed alternatives can be cheaper once setup labor is counted[7]
  • Autonomy cuts both ways — Less human-in-the-loop control than approval-first tools like Cline; review happens at the PR, not at each action
  • Cloud free tier is capped — Individual plan limits you to 10 conversations per day[6]

Pricing & Licensing

TierPriceIncludes
Open Source (Local)FreeFull agent, web GUI, terminal UI, CLI, git integrations, community support[6]
Individual (Cloud)Free to startHosted access, API, Jira/Slack integrations, 10 conversations/day; BYOK or OpenHands provider at cost with no markup[6]
EnterpriseCustomSaaS or self-hosted in your VPC, SAML/SSO, unlimited concurrent conversations, Large Codebase SDK, priority support[6]

Licensing model: MIT for the core platform and Docker images; the enterprise/ directory is source-available and requires a paid license to run beyond one month.[1]

Hidden costs: Inference. Self-hosted tasks run roughly $0.15–$0.60 each in API fees with frontier models, and setup/maintenance time is real overhead versus managed alternatives.[7]


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
DevinClosed, subscription-priced managed service; OpenHands is its open-source, self-hostable counterpart
ClineCline is human-in-the-loop with per-action approvals in VS Code; OpenHands is autonomy-first with review at the PR
AiderAider is a supervised terminal pair programmer in maintenance mode; OpenHands targets full task delegation and ships daily
OpenCodeOpenCode is a terminal-first open agent; OpenHands adds GUI, cloud, sandboxed runtime, and an agent SDK
Claude CodeFirst-party Anthropic CLI agent; OpenHands is provider-neutral and adds hosted/self-hosted GUI surfaces
Jules / Codex (cloud)First-party cloud agents from Google/OpenAI; OpenHands Cloud is the open, model-agnostic equivalent

When to Choose OpenHands Over Alternatives

  • Choose OpenHands when: You want to delegate whole tickets to an autonomous agent, need self-hosting or an open MIT core, or want to build custom agents on an SDK
  • Choose Cline when: You want agentic AI inside VS Code with explicit approval of every action
  • Choose Aider when: You prefer a minimal, stable terminal pair-programming workflow with git-native commits
  • Choose OpenCode when: You want an actively developed open-source terminal agent without the platform footprint
  • Choose Devin when: You want a fully managed autonomous agent and will pay for zero setup

Ideal Customer Profile

Best fit:

  • Teams that want to delegate bug tickets, migrations, and maintenance work to background agents
  • Organizations requiring self-hosted or VPC deployment with full code/data ownership
  • Platform engineers building custom agents on the Python SDK
  • Enterprises wanting an open, model-agnostic alternative to Devin
  • Developers comfortable with Docker and LLM API configuration

Poor fit:

  • Developers who want per-action approval and tight supervision (see Cline)
  • Frontend-heavy teams — visual/UI work is a documented weak spot[7]
  • Anyone expecting a zero-setup, download-and-run experience
  • Budget-constrained users on weak models — the agent needs frontier LLMs to perform

Viability Assessment

FactorAssessment
Financial HealthStrong — $18.8M Series A (November 2025) led by Madrona, atop a $5M seed[2]
Market PositionLeader — most-starred open coding agent at 76K+ stars; the open-source reference point for autonomous agents[1]
Innovation PaceRapid — daily commits, SDK rearchitecture, cloud and enterprise surfaces all shipped within the past year[1]
Community/EcosystemStrong — 9,700+ forks, active Slack community, academic benchmark infrastructure[1]
Long-term OutlookPositive — clear open-core revenue model; main risk is first-party cloud agents (Devin, Jules, Codex) commoditizing the category

OpenHands has converted academic credibility (OpenDevin, SWE-bench) into the strongest community position in open autonomous coding agents, and the November 2025 raise plus rebrand signal a deliberate push from research project to enterprise platform. The competitive question is whether an open, model-agnostic platform can out-execute the first-party cloud agents now shipping from every major lab.


Bottom Line

OpenHands is the open-source standard-bearer for fully autonomous coding agents — the tool to reach for when you want to hand an agent a ticket and review a PR, not babysit an editor session. The 76K-star community, MIT core, model-agnostic architecture, and every-surface strategy (SDK, CLI, GUI, cloud, VPC) make it the most complete open platform in the category, but it demands technical comfort: Docker-heavy setup, frontier-model costs, and tolerance for the occasional agent loop.[7]

Recommended for: Teams delegating real engineering tasks (bug fixes, migrations, triage) to autonomous agents, organizations needing self-hosted deployment, and builders who want an SDK rather than a finished product.

Not recommended for: Developers who want supervised, approval-per-action workflows, frontend-heavy work, or a frictionless managed experience out of the box.

Outlook: With fresh capital, daily shipping velocity, and the open-core enterprise model in place, OpenHands is positioned to remain the default open alternative to Devin-class agents. Watch whether the cloud business converts community scale into revenue before first-party agents from the model labs absorb the autonomous-agent category.


Research by Ry Walker Research • methodology