OpenHands in 2026: From CLI to the Agent Canvas Platform
The OpenHands CLI is officially deprecated in 2026. Meet Agent Canvas, learn the right way to use it, and see how it pairs with Hermes Agent.
TL;DR
OpenHands dropped its pip CLI and pivoted to Agent Canvas, a canvas-based platform installable via npm or a safer Docker sandbox scoped by PROJECTS_PATH. The SDK adds security risk levels, LiteLLM provider flexibility, and named LLM configs for cost tiering. Pair it with Hermes Agent as an orchestrator that verifies diffs before anything gets merged.
If you try to install OpenHands with pip install openhands the way older tutorials show, what appears is not an installation guide but a notice: "This project is no longer actively maintained." That does not mean the project is dead — OpenHands is in fact extremely active. What died is its old form.
Throughout 2026, OpenHands changed completely: from a simple CLI tool into a platform called Agent Canvas. For anyone following AI coding agents, most information circulating from 2024–2025 is now outdated. This article covers what changed, the right way to use it today, and where it stands if you also run Hermes Agent.
What changed in 2026
The OpenHands GitHub repository now holds roughly 86.3 thousand stars with a beta status, and Agent Canvas 1.16.0 shipped in August 2026, marking the pivot from a CLI to a canvas-based interface. Commits were still landing as of September 2026.
This direction is not community speculation. Official issue number 14374 in the OpenHands repo states it explicitly: "Agent Canvas will be the main interface for OpenHands." This is the product's main direction, not a side feature.
The practical consequence: the V1 CLI, once installed via pip install openhands, is officially deprecated. The PyPI package is marked as no longer actively maintained and points users to Agent Canvas. If you still use the old CLI, that is your signal to migrate.
The right way to use it today
There are two installation paths for Agent Canvas, and the difference matters for security.
Path one: global npm. Run npm install -g @openhands/agent-canvas, then invoke agent-canvas from your terminal. This mode has no sandbox, so the agent gets full filesystem access. It fits a personal work machine, not shared environments.
Path two: Docker. Run the ghcr.io/openhands/agent-canvas image. In Docker mode the agent can only see folders designated through the PROJECTS_PATH variable. This is a true sandbox, and the safer choice for work touching many files or sensitive data.
Deeper needs, such as building your own agent or wiring it into a custom pipeline, are served by the OpenHands SDK in the separate software-agent-sdk repository. Installation is two packages at once, pip install openhands-sdk openhands-tools, and both must stay on the same version; upgrading in a single command keeps them from drifting.
For language models, OpenHands uses LiteLLM as its adapter, so any provider works. OpenHands' own cloud models use the openhands/ prefix. A rarely discussed feature: config.toml supports named LLM configs, for example [llm.repo-explorer], which inherits everything from the default [llm] section. That opens the door to cost tiering: a cheap model for repository exploration, a strong model only for code generation.
For local models, Qwen3.6-35B-A3B is officially supported through LM Studio, Ollama, vLLM, or SGLang, with a minimum context of roughly 22 thousand tokens. One common trap: LM Studio by default only listens on the local interface, unreachable from Docker containers. Enable "Serve on Local Network" so the model server is reachable from your local network.
Often overlooked: security and limits
In the SDK, every agent action is assigned a risk level by a SecurityAnalyzer: LOW, MEDIUM, HIGH, or UNKNOWN. That classification decides when ConfirmationPolicy requires manual approval. Disabling it for speed means discarding the last safeguard before an agent deletes files or writes unverified code.
Another explicit limit: the public runtime at runtime.all-hands.dev is officially for benchmark evaluation only and is not allowed for production. Production needs are directed to OpenHands Cloud Workspace.
The subscription login feature, which lets a ChatGPT Plus/Pro subscription drive Codex models without API costs, also deserves testing first. There is a documented bug: the gpt-5.1-codex-max profile fails on the very first model call when used with the default agent (issue #15528). Keep a fallback to direct API keys in case this blocks your workflow.
OpenHands + Hermes Agent: not rivals
OpenHands' official position today is an "outer loop" platform: an agent picks up a work ticket, completes it, and hands back a pull request for review. Hermes Agent sits on a different side: a personal orchestrator with persistent memory, skills, cron scheduling, and citation-backed research. The question is not which one is better but how the two complement each other.
The integration pattern is concrete. Hermes can launch an Agent Canvas Docker sandbox for work that needs isolation: large refactors, cross-package dependency bumps, or long-running automations, with the folder scope locked through PROJECTS_PATH. The result comes back to Hermes for verification: the diff is reviewed, tests are run, and only then is the work accepted or discarded. In this pattern OpenHands never pushes to a repository; Hermes acts as the final gate.
A setup that saves money: any local OpenAI-compatible model proxy can be registered as the LLM endpoint in OpenHands' config. One model backend then serves both agents without duplicating providers.
For a map of coding agents on the Hermes side, including how OpenHands compares with Claude Code and OpenCode as delegation skills, read the article Review of 3 Coding Agent Skills in Hermes Agent.
Who needs what
Use Agent Canvas when you need an isolated sandbox, long-running automation, or a workflow that ends in a pull request for review. Hermes Agent alone is enough when daily work is interactive: debugging, code exploration, research with long context, or orchestrating many tools. And when you need both, a single local model proxy is enough to connect them.
Sources
1. OpenHands/OpenHands repository
2. Issue #14374 "Agent Canvas will be the main interface for OpenHands"
3. OpenHands CLI V1 deprecated
4. Agent Canvas installation (npm & Docker) — OpenHands repository README
6. Custom LLM Configs via LiteLLM