GILD verdict
Choose OpenHands if your team wants to own its coding agent, from the code it runs to the model contracts it bills against.
Budget for a well-configured sandbox, a strong model with a spending cap, and repository setup that lets the agent check its own work.
OpenHands
AI Coding Agent by All Hands AI
+15 MixedFit and limits
Who it fits
- Security or data rules require the agent, its model calls, and its sandboxes inside your own infrastructure. OpenHands Enterprise runs on Kubernetes in your VPC, and the open-source Agent Server runs on any VM, Docker host, or Kubernetes cluster you choose.
- You want model calls billed under your own contracts with Anthropic, OpenAI, AWS Bedrock, Azure, or Google Vertex AI, or want open-weight models. OpenHands connects to any model LiteLLM supports, and the at-cost OpenHands provider covers anyone without a key.
- Engineers want to inspect or extend the agent. Agent Canvas, the Software Agent SDK, and the CLI are MIT-licensed, and the SDK builds custom agents and GitHub workflows on the same engine.
- You want one control center for several agents. Agent Canvas runs the OpenHands agent and can also drive Claude Code, Codex, and Gemini CLI through the Agent Client Protocol.
Where it falls short
- Results depend on the model. As of Oct 25, 2026, All Hands AI's own index ranges from 83.8% on SWE-bench Verified with Claude Opus 4.8 to 34.2% with Nemotron-3-Nano. Its docs warn that open-weight and local models vary widely in tool-use reliability.
- On the independent Terminal-Bench 2.0 leaderboard, the OpenHands harness trails the leaders. With Claude Opus 4.5 it scored 51.9% on January 4, 2026, and Factory's Droid scored 63.1% on the same model.
- Model spend sits outside the OpenHands plan and the docs warn the agent issues many prompts, so a team without budgets has no ceiling on cost.
- Running it yourself means owning the sandbox: the process sandbox has no isolation, and broad mounts, privileged mode, host networking, or access to the Docker socket weaken a container.
- The local open-source setup is built for one user, with no built-in authentication or multi-tenant isolation. Shared team use needs Teams at US$500 a month for up to 10 users, or Enterprise by quote.
Implementation reality
Choose the deployment first
Local Agent Canvas suits one engineer. A self-hosted Agent Server on a VM, Docker, or Kubernetes serves a team you trust with one backend. OpenHands Cloud runs managed sandboxes with organizations, budgets, and integrations, and Enterprise moves all of it onto your Kubernetes cluster with SAML SSO, RBAC, a LiteLLM gateway, and analytics.
Pick the sandbox deliberately
Docker is the default and recommended sandbox. The process sandbox runs the agent with your user's access to files and commands, so keep it for controlled environments. Mount only the project directories the agent needs, since it can change or delete anything mounted read-write.
Set a model and a ceiling
Start from the OpenHands Index: Claude Opus 4.8 averages 71.9, GPT-5.5 65.9, and Gemini 3.5 Flash 62.6 across five benchmarks, with GLM-5.1 at 58.2 as the strongest open-weight entry. The index also prices each SWE-bench Verified task: US$0.75 on Claude Opus 4.8, US$1.52 on GPT-5.5, and US$1.54 on GLM-5.1. Set a maximum budget per conversation, and in Cloud organizations a monthly organization budget with alerts at 80, 90, and 100 percent.
Scope repository access
The OpenHands Cloud GitHub app requests 8-hour tokens with read and write access to Actions, contents, issues, pull requests, webhooks, and workflows on the repositories you select. Grant it only the repositories in the pilot, and keep branch protection so a human approves every merge.
Give the agent a way to check itself
A setup script in .openhands/setup.sh prepares dependencies each time the agent starts. An AGENTS.md file carries build and test commands. A Stop hook can block the agent from finishing until your tests or linters pass. On OpenHands Cloud, a team running setup scripts in 9 production repositories reports that the sandbox home directory resets between turns. Every session then reinstalls dependencies from scratch (GitHub issue #18942, October 25, 2026).
What to measure in your pilot
- Track model spend per merged pull request for each allowed model over two weeks; drop any model that costs more than double its OpenHands Index price per task.
- Time setup.sh on every pilot repository; where it runs longer than 3 minutes per conversation, bake dependencies into a custom sandbox image.
- GILD found no team account of splitting work across Agent Canvas agents outside All Hands AI's blog. Send one task type to Codex for two weeks and keep whichever costs less per merge.
Sources
- Official documentation
- 22
- Public source
- 5
- OpenHands pricing
- OpenHands Enterprise
- OpenHands on GitHub: Agent Canvas (MIT)
- OpenHands on GitHub: LICENSE
- OpenHands docs: Introduction
- OpenHands docs: FAQs
- OpenHands docs: LLM overview and model recommendations
- OpenHands docs: OpenHands LLM provider and pricing
- OpenHands docs: Sandbox overview
- OpenHands docs: Docker sandbox
- OpenHands docs: Process sandbox
- OpenHands docs: Agent Canvas architecture
- OpenHands docs: ACP agents in Agent Canvas
- OpenHands docs: GitHub integration
- OpenHands docs: Application settings
- OpenHands docs: Organization budgets
- OpenHands docs: Repository customization
- OpenHands docs: General skills and AGENTS.md
- OpenHands docs: OpenHands Enterprise
- OpenHands docs: Enterprise vs. open source
- OpenHands docs: Enterprise architecture
- OpenHands Index results repository
- Snorkel AI: Terminal-Bench 2.0 leaderboard, Stanford University and Laude Institute (as of Oct 25, 2026)
- OpenHands Index: Claude Opus 4.8 scores and cost per task (submitted Jun 2, 2026)
- OpenHands Index: GPT-5.5 scores and cost per task (submitted May 8, 2026)
- OpenHands Index: GLM-5.1 scores and cost per task (submitted Apr 24, 2026)
- GitHub issue #18942: dependency-cache persistence for .openhands/setup.sh across sandbox HOME recycling (Oct 25, 2026)