OpenHands, the open-source AI development project with over 83,000 GitHub stars, released Agent Canvas 1.10.0 on August 5, 2026, then followed with 1.11.0 and 1.12.0 on August 7, 2026. On its own, each release looks like a small polish pass — a new default model here, an export button there, a cost figure in the activity log. But read together, the August release train continues a clear push to make self-hosted coding agents feel more reliable, more transparent, and easier for newcomers to control.
If you are new to OpenHands — or you have been wondering whether it is worth trying over Claude Code or Codex — this article explains what the project is, what the 1.10–1.12 releases actually changed, and why the small updates matter more than a version bump suggests.
What OpenHands Is (If You Are New)
OpenHands now brands itself as Agent Canvas: a self-hosted “control center” for AI coding agents. Instead of just being one more terminal agent, it is a web app that runs on your machine (or a server) and lets you start conversations, review what the agent did, and set up recurring automations.
Key facts:
- It runs the open-source OpenHands agent out of the box, but can also drive Claude Code, Codex, Gemini, or any agent that speaks ACP (Agent-Client Protocol).
- It is self-hosted by default — your code and prompts stay on your hardware unless you opt into OpenHands Cloud.
- You can run the agent locally, in a Docker container, on a VM, or on a cloud server, and switch backends from the same interface.
For beginners, the practical draw is one UI that lets you watch, pause, and automate an agent instead of trusting it blindly against your whole filesystem.
What’s Actually New in the Recent Release Cycle
Version numbers come fast (1.8 → 1.9 → 1.10 within a week), so here are the changes that matter for a newcomer, in plain language.
1. Persistent agent memory (the headline feature)
In the 1.7.0 cycle, OpenHands added a persistent agent memory toggle. Instead of forgetting everything between chats, the agent can keep a short-term memory across sessions. For a beginner this is a genuinely useful quality-of-life change: you no longer have to re-explain your project layout, your naming conventions, or your goals every single conversation.
2. You can finally see what the agent is doing, live
1.9.0 added live agent activity in the chat view, and 1.10.0 adds an activity log export. This is a small thing with a big trust benefit. When a coding agent runs with full filesystem access, “trust but verify” is the only sane approach. Being able to see the agent’s steps as they happen — and export a log later for review — gives you a transparency layer you previously did not get in self-hosted setups.
3. Automations get a proper dashboard
OpenHands lets you set up agents that run on a schedule or in response to events (a Slack message, a new GitHub issue, a webhook). 1.10.0 ships a featured automations landing dashboard and adds manifest-driven sub-pages, so the whole flow of creating, browsing, and enabling automations is closer to a proper product and less like fiddling with config.
4. MCP server controls and creds stay put
MCP (Model Context Protocol) servers are how agents get extra tools. OpenHands already lets you enable/disable an installed MCP server from its card; the latest release also fixes a real foot-gun: MCP credentials are preserved during the editor’s mutations, so tweaking a server’s config no longer silently drops a secret-like API key you plugged in.
5. A new out-of-the-box default model
1.10.0 sets the Canvas default to GLM 5.2. If you do not pick a model, this is what your agent uses. Model choice still comes per project — the default matters mainly because a sensible, cheaper model right off the bat lowers the “wow, this cost me money” surprise for first-time users.
6. Skills surfaces
OpenHands keeps its skills filter with a faceted rail in 1.10.0, and earlier fixes stopped disabled skills from being loaded into the agent’s context. Put simply, the project is cleaning up context/bloat so your agent reads less irrelevant noise and works on exactly what you asked.
What Landed After 1.10: 1.11.0 and 1.12.0 (August 7)
Two days after 1.10.0, OpenHands shipped 1.11.0 and 1.12.0. Both are small, but two changes are worth knowing about even if you are brand new.
1. You can now see what each session cost (1.11.0)
The biggest addition for beginners is per-run LLM cost in the Activity Log and its exports. Previously, answering “how much did this agent session actually cost me?” meant digging through provider dashboards. Now the figure shows up inside OpenHands itself, next to the session, and travels with the exported log.
For people new to self-hosted agents this is a genuinely useful trust feature: it removes the surprise-bill anxiety that makes newcomers afraid to let an agent work (see our guide on silent quota and model burn). OpenHands also changed its chat-view navigation to show the control overview and layout.
2. Automations get tags and child conversations (1.11.0)
If you set up several automations (a Slack-triggered agent, a daily issue sweeper, a webhook responder), 1.11.0 adds tag filters for automations plus tag chips on conversations, so you can group and find things instead of scrolling. It also adds a typed agent action that can launch a child conversation — the agent itself can spawn a local or cloud conversation and come back with the result. That is the building block for multi-agent workflows (“spawn a second agent to review my PR”), and it is now available as an explicit, typed action rather than a prompt workaround.
3. The desktop app is now “OpenHands Agent Canvas” (1.11.0)
OpenHands rebranded its desktop app to Agent Canvas to match the web control center, and improved the version-update UI. If you were confused about whether the desktop app and the web app were the same product, they now share the same name and direction.
4. Free and paid models are clearly labeled (1.12.0)
The headline of 1.12.0 is a single, product-level fix: clarifying which OpenHands model endpoints are free. First-time users often pick a model without knowing whether it silently costs money or runs on a free tier. 1.12.0 labels those endpoints so “free” actually means free — a small anti-bill-shock change in the same spirit as the default-model work in 1.10.0.
How To Try It (Three Ways)
Option 1 — fastest (no sandbox), thinking beware: runs the agent directly on your machine with full filesystem access.
npm install -g @openhands/agent-canvas
agent-canvasOption 2 — safer, with a Docker sandbox: the agent only sees the folder you give it.
export PROJECTS_PATH="$HOME/projects" # folders the agent may touch
mkdir -p "$PROJECTS_PATH" "$HOME/.openhands"
docker run -it --rm -p 8000:8000 \
-v "$HOME/.openhands:/home/openhands/.openhands" \
-v "${PROJECTS_PATH}:/projects" \
ghcr.io/openhands/agent-canvas:1.12.0Option 3 — from source: clone OpenHands/OpenHands, run npm install then npm run dev, and open http://localhost:8000.
At a Glance
| Feature | Where it landed | Why it matters for you |
|---|---|---|
| Persistent agent memory | 1.7.0 | Agent remembers project context between chats |
| Live activity + activity-log export | 1.9 / 1.10 | You can see/verify what the agent is doing |
| Automations dashboard | 1.10.0 | Recurring agents are easier to set up and manage |
| MCP credentials preserved | 1.10.0 | Editing config no longer drops secrets |
| GLM 5.2 default model | 1.10.0 | Cheaper, safer default out of the box |
| Skills context cleanup | 1.8–1.10 | Less token noise, sharper behavior |
| Per-run LLM cost in activity log | 1.11.0 | You can see what each agent session costs |
| Automation tags + child conversations | 1.11.0 | Find automations fast; multi-agent building block |
| Desktop app = Agent Canvas | 1.11.0 | One name for web and desktop |
| Free endpoints clearly labeled | 1.12.0 | No surprise bills on “free” models |
Bottom Line
1.10.0 is not an earth-shattering release on its own — most commits are UI and internal plumbing — and neither are the 1.11.0 and 1.12.0 follow-ups. But read together with the persistent-memory work and the live activity view, the direction is clear: OpenHands is spending its effort on visible control, verified behavior, and lower cost of entry for self-hosted agents. Per-run cost tracking (1.11.0) and clear free-model labels (1.12.0) go straight at the two anxieties that scare newcomers away from letting an agent work: “is it wasting my money?” and “will I get a surprise bill?” For a beginner, that is exactly the right place to spend the effort.
Start with the Docker sandbox option so the agent lives in a box you can throw away, enable persistent memory once you trust it, use the activity log to review what it did — and check the per-run cost figure after your first long session so the numbers never surprise you. That pattern keeps the power of a coding agent without giving an autonomous process free rein over your machine or your wallet.
New here? Start reading next:
- Your AI Agents Config Directory Is Now the Most Dangerous Place on Your Machine
- Cline v4.1.0: Five Critical Agent Fixes and the SDK Migration That Stops AI From Going Rogue