OpenHands Agent Canvas

What Agent Canvas is

OpenHands spent three years building one of the strongest autonomous coding agents - the project behind much of the early SWE-bench conversation, formerly All Hands AI, independent and $18.8M Series A funded (Madrona, November 2025). Agent Canvas, announced June 16, 2026, is the company’s bet on what comes after one-shot prompting: a self-hosted control center where coding agents run continuously against real repos.

The frontend speaks to an Agent Server (a Python REST API that runs multiple agents on one host), an automation server for schedules and webhooks, and any ACP-compatible harness. It is MIT licensed and self-hostable from a laptop up.

Install and sandbox

Three install paths. The no-sandbox path (npm install -g @openhands/agent-canvas, then agent-canvas; Node 22.12+ and uv required) gives the agents full filesystem access - the docs say so plainly. The Docker sandbox path runs ghcr.io/openhands/agent-canvas:1.16.0 and only exposes a directory you choose: agents can touch what is under your PROJECTS_PATH mount and their ~/.openhands state, and nothing else. The UI lands on localhost:8000. The docs flag extra hardening steps before exposing anything on a network.

What automations add

A workspace plus a scheduler is the real product change. A workflow defined once can run as a cron job - weekly repo summaries, dependency checks, docs-drift checks - or fire on events from Slack, GitHub, and Linear: a PR opens, a review request lands, a channel gets a message. Each run has history you can inspect. OpenHands positions this as the operational layer: durable, trigger-based workflows instead of one-off terminal sessions.

Harnesses and models

Agent Canvas is not tied to its own agent. Over the Agent Client Protocol (the same protocol Zed pushes) it drives Claude Code, Codex, and Gemini CLI against your existing subscriptions - Claude Max, Codex Pro - and the OpenHands agent stands beside them. LLM profiles let you assign a cheaper model to scheduled background runs and switch models mid-conversation. The company ships NVIDIA Nemotron support and an AMD Instinct deployment guide that pairs Qwen3-Coder with vLLM, so open-weight models on your own GPU are a first-class path, not an afterthought.

Where it fits for local AI

The pitch - your agents on your machine, your model choice, your file access boundary - is the local-AI argument applied to coding agents. Run the Docker variant on a Mac Studio or a GPU box, point the harnesses at open-weight endpoints, keep the source code and the automation history under your roof, and the Enterprise control plane stays optional.

Sources: Agent Canvas announcement, GitHub, docs.