alert Hi Everyone! We'll have our next Community Call on August 6th at 12pm ET. We'll send invites via Luma in advance of that, and also post links to signup to join the community call early next week!

Hi guys, I've tried multiple times to register on OpenHands Cloud on Safari, Chrome, and Firefox using GitHub and it never worked. It keeps saying that the email verification is sent but I received nothing.

Also tried GitLab and BitBucket

😹 Opus 5. On Friday evening, because of course.

🙌 Agent Canvas 1.6.1 is out the door! Some great improvements to see when you update: • MCP server cards now show live connection health with actionable status indicators for quick troubleshooting. • Plugin detail modal displays the full contents of a plugin — skills, commands, and MCP configs — before install. • Update-availability card added to Settings to show users when a newer Agent Canvas version is ready. • Cloud-aware Skills and Plugins navigation with a new Integrations settings link. • Title generation profile preference lets users pick which LLM generates conversation titles. • Streaming reliability improvements: final events render authoritatively and reconnect side-effects are deduplicated. • Cloud backend reconnect modal no longer locks the UI. • Secret values are preserved when editing an existing secret. • LLM Profiles remain reachable in Settings while an ACP agent is active. • Paginated workspace folder listings now load correctly. • Dev services on Windows spawn without a shell, fixing local development startup. Give a shout in #C06QT0AGY4W if you have any feature requests, ideas, or run into any issues

Hi everyone! I'm working on an AI research project that experiments with *algorithm synthesis*—trying to discover or combine solution strategies automatically rather than manually designing them. As one experiment, the system generated a proposed Rubik's Cube solving method called DR-OPT (Diminishing Returns Optimal). Since I'm not actually a speedcuber, I can't confidently judge whether the synthesis is genuinely novel, practical, or simply rediscovering existing ideas. I'd really appreciate feedback from experienced cubers. Specifically: • Does this resemble an existing method? • Are there incorrect assumptions or impossible claims? • Does the overall reasoning make sense? • If you were to critique it, what would you immediately question? Document: github.com/tamzrod/kde/blob/…/DR-OPT-MANUAL.md I'm more interested in understanding whether the synthesis process produced something meaningful than proving that it's the "best" method. Any criticism is welcome.

I tried OH at the beginning of the year via docker on my proxmox server in a debian vm. I gave up in frustration after a few days trying to get the sandbox to initialize. I remember here,on this very slack channel, some of the OH devs said it was a known issue with running OH remotely as it usually is installed on and used on the same computer. What I'm trying to achieve is running it on a vm in proxmox then connect to it via my office computers browser and use it. Has the bug or issue that prevented this before been resolved?

So, I got OH installed via docker last night. I'm having a issue, I think. I provided some screen shots. You can see the welcome chat or the create a basic webpage that we get after install that the model is reporting deepseek. I did connect to my deep seek account and was playing around with some code. Then I wanted to see if I could get opencode zen platform to work so I could use their free tier big pickle model. I setup the opencode zen account, then I switch the model via the selection option in the chat to the opencode provider from the deepseek one. I thought it was working but realized my deep seek usage was still growing. So, I asked it what model it was and to my surprise it said deep seek. While the chat box said opencode the left pane still showed deepseek, so I refreshed and it changed but it still reports being deep seek. I started a fresh chat and interacted with opencode right from the start to make sure it's working and it seems like it is..it reports being opencode zen. Is there a bug here, shouldn't I be able to switch models and providers in the same chat/project. Also, how do you get your files? I was playing with getting it to compile tiddlywiki for cloudfare and it said it finished and had the zip file but I can't find it anywhere in the file tree on the right.

So, I went back to the original chat that was started with deepseek and asked it if it was big pickle ...just because not only did it say it was deep seek but if you look in the left hand window you can see where it changed it's profile to deep seek while the selection in the chat says opencode. It changed while I asked it that question.

Hey everyone - I’ve been working on a fault-injection reliability benchmark for OpenHands agent runs: RFC: github.com/OpenHands/benchmarks/issues/764 Implementation PR: github.com/OpenHands/benchmarks/pull/765 The benchmark runs matched baseline/faulted tasks under deterministic sandbox restarts, lost dispatch responses, SIGKILL mid-tool-call, and network partitions. It produces inspectable evidence for: • whether the run resumes and completes • whether replay duplicates an irreversible effect • how much recovery overhead it adds The benchmark core, graders, SDK event bridge, CLI, docs, and tests are implemented. Runtime-specific fault operations remain behind explicit adapters. I’m looking for guidance on two decisions: 1. Is `OpenHands/benchmarks`the right long-term home, or should part of this live in the SDK? 2. Which real runtime/SWE-bench adapter would be the most useful first integration? If anyone owns benchmark infrastructure or runtime recovery, I’d really appreciate a pointer or review. I’m happy to adjust the scope around what would be easiest to land.

Hey team! I want to introduce you to a tool that I use with OpenHands. Its called Maestro and it intelligently delegates tasks to different models and harnesses depending on complexity and size of the task. The architecture is something like this:

Maestro

            Task arrives
                 │
                 ▼
          Analyze complexity
                 │
      ┌──────────┼──────────┐
      ▼          ▼          ▼
 Claude Code  OpenHands   Cursor Agent
      │          │          │
      └──────────┼──────────┘
                 ▼
            Validation
                 ▼
        Architecture review
                 ▼
         Benchmark result
                 ▼
       Learn for next time

Do you have any tools you use with OpenHands that you recommend? I'm always looking for new/better ways of working 😄

I'm trying open hands again since you guys fixed the sandbox initializing issue for running on remote machines. I'm still having a few issues. To start, I'm running the -latest in docker so I believe it's 1.8.0 released yesterday. It's running in a lxc linux container on my proxmox machine. So the backend and frontend are on the same machine. I'm connecting to it remotely on the same LAN from my desktop computer. It keeps loosing connection to the backend, I get a notification in the upper right corner that it lost connection and it just hangs, sometimes I have to refresh and sometimes I have to exit the firefox / brave tab and connect to it again. Other times it will go to this screen I'm attaching for reference. If I refresh or restart a browser connection it seems like it does continue on. The linux container isn't cpu or memory bound. I've looked at it's resources in the proxmox summery and it's fine and I can still ssh to the container.

Another issue is If I start a new conversation and before I type anything I use the switcher to select the LLM model I want...It will start with the default even though I asked it to start with my selection.

hey i see on openhands cloud you guys have own LLM

Hi all! For a limited time, we’re offering deepseek-v4-flash for free on the OpenHands Cloud! You can log in (https://app.all-hands.dev/canvas) and add an LM profile like follows to try it out.

Does the open hands agent canvas self hosted version plan on getting github integrated like the cloud version?

I wanted to run the Openhands agent sandbox-server as remote in Daytona cloud 's container, so can i know the recommended specs for the container (i mean ram, vcpu, storage) ??

We need vpn capabilities, one to securely route traffic but also to pick different endpoints globally. Sometimes the endpoints in NA are overloaded or we're hit with higher fees for using LLM endpoints coming from a certain geographical area. I use nord and proton for routing my normal desktop browser traffic as needed. It would be nice to have this in OH. You already have a built in browswer to the subsystem is there...maybe a simple plugin engine would work. Or, just create a vpn section in the settings and let the user setup tailscale, wireguard etc. Also, have the ability to create several configs so a user can easy switch from one vpn service or endpoint to another in the UI like choosing a llm.

Hello, I'm new here. As an AI researcher, I learned about this tool and would appreciate guidance on where to find the technical architectural documentation for reading.

👋 Join me on in the Community Call in ~5 minutes. We'll be talking about the updates to the SDK and OpenHands in the past month. We'll also be showing off how we have agent canvas automations running code review, slackbots, triage and more for the OpenHands repos. OpenHands Community Meeting Thursday, August 6 · 12:00 – 1:00pm Time zone: America/New_York Google Meet joining info Video call link: meet.google.com/xvv-bdgs-snt

Hi all! I made a video of how I manage my daily work with agents, including reading my communications on slack/email, and managing my to do list: Sharing in case it’s interesting 🙂

I'm thinking of moving from the docker version to a native install in a vm, I'm tired of the mapping issues. Is the native version recommended and is it easy to update? One thing about docker is it's so easy to just do a pull. From you guys that use OH daily do you prefer docker or native?

Hello, everyone! I’ve spent the past few years building AI and full-stack systems, recently working on an AI software maintenance platform that combined coding agents, repository analysis, isolated execution, automated testing, and GitHub workflows to turn issues into reviewed pull requests. I’m especially interested in agentic software engineering and how AI can handle more of the development lifecycle, and I’m always open to new opportunities and collaborations.

I think the interesting shift with coding agents is moving from “generate this function” to delegating complete engineering tasks. But autonomy only becomes useful when the agent can inspect the codebase, understand existing architecture, make changes safely, run meaningful tests, and provide evidence for its decisions. The challenge isn't just making agents capable of coding—it’s making their work predictable, verifiable, and easy for engineers to review.

This message was deleted.

Open request: Can we get a plugin in engine so plugins can be made. Perhaps a repo like other apps where one can brows and see what's available and then install it? First thing I'd ask a plugin for is something like a decent IDE. It'd like to follow along with the agent and or even interact and modify the code by hand if necessary. I like how antigravity worked before Google nerfed it. OH is almost a drop in replacement minus the ide. Perhpas this project could be used as inspiration for a integrated web IDE? theia-ide.org

The OpenHands Index hasn't been updated in a while. Will new models be added, please?

One common failure mode I have seen recently is that the agent will implement something in the wrong repository. For instance we keep getting new implementations of API access to endpoints in the OpenHands repo despite the fact that this should be in typescript client. We could consider restructuring repos but as a first step maybe we can improve our agents.md and review guidelines about things that should not be implemented in particular repose but a different repo

hi OH, i found this and verified it Issue: Automation master API key exposed to agent sandbox Our automation backend's master API key (

OPENHANDS_AUTOMATION_API_KEY

) is currently injected as a plain environment variable into the AI agent's sandbox, the same environment where it executes arbitrary shell commands. This violates least-privilege: the sandbox is a lower-trust context (commands can be influenced by untrusted input via prompt injection, or dumped accidentally via

env

, debug output, or crash logs), so any leak there exposes a full-privilege credential capable of controlling the entire automation backend. We should either stop seeding this key into the agent's environment (check injection point in

dev-with-automation.mjs

) or replace it with a short-lived, narrowly-scoped token so a leak has limited blast radius.

I need some help with getting Agent-Canvas to work. I installed it in a Docker container running on Unraid, and I made a connection with the backend. However, i can't make a connection with the Ollama container running also on Unraid, where my models are stored. I do a test i get an error like this: litellm.APIConnectionError: OllamaException - Cannot connect to host host. docker.internal:11434 ssl:<ssl.SSLContext object at 0x14b6591dd130> [Connect call failed ('192.168.X.X', 11434)] I don't understand this cause i don't know why it looking for an SSL connection while I just want it to connect internally to my Ollama container. I also don't know why I failed to connect to a different IP address, which I did not configure. The IP address it is mentioned as failed is the IP address of my reverse proxy. I don't know why it connects to that ip address with the port number of the Ollama container.