Summary
- The OpenClaw Foundation released version 2.0, its largest update yet, incorporating more than 16,000 pull requests
- The new version auto-detects existing API keys, ChatGPT/Claude subscriptions, and local models, cutting down the initial setup process
- Shared Cloud Sessions let multiple users work on the same task together, while Crabbox adds support for spinning up temporary cloud instances
The OpenClaw Foundation has released version 2.0 of OpenClaw, its open-source AI agent platform. The release folds in more than 16,000 pull requests, making it the largest update the foundation has ever shipped. At the center of it are two changes: an automated setup process that replaces the complicated configuration users used to face on first run, and new shared cloud sessions that let multiple people work on the same task together.
In plain terms, OpenClaw is an open-source AI agent that creates files, writes code, and handles tasks on your behalf inside a computer. It's built so sessions can run wherever you want — on a local machine, on another device you've connected yourself, or on a rented cloud server.
What OpenClaw is
OpenClaw is an open-source AI agent that creates files, writes code, and handles tasks inside a computer. Its full source is available in the openclaw/openclaw repository on GitHub, so anyone can download it and run it locally. Our tracking system has picked up movement around this project before, but this is the first time we're covering it as a full story.
Setup just got a lot shorter
The biggest change is what happens when you first launch the tool. Previously, users had to manually enter things like API keys, subscription details, and local model paths. Starting with 2.0, the software automatically detects any ChatGPT or Claude subscriptions, API keys, and local models already present on the machine. Everything else can be sorted out later through conversation with the bot. The chat window opens immediately on launch, letting you continue setup, check on running tasks, or watch progress unfold in real time.
Inside the chat window there's also a new status bar called the "Session Rail," which shows a running session's rating, plan progress, and pull request status at a glance. A companion chat panel has been added too, letting you ask questions about the current session without interrupting the agent. According to the documentation, the gateway pulls in a limited snapshot of the session and answers using a separate utility model.
Cloud sessions multiple people can share
The new Shared Cloud Sessions feature lets multiple users join and work on the same task together. It's built so you can pull teammates into an ongoing task, or hand off work along with its context. The OpenClaw Foundation says it's already using this feature internally to build OpenClaw itself.
How to try it
- Where you start — Launching OpenClaw for the first time opens straight into a chat window instead of a setup screen.
- Automatic connection — Answering a few questions in the chat lets OpenClaw automatically pick up any API keys, subscriptions, and local models already on your machine. Anything still missing can be filled in later through conversation.
- Checking sessions — The Session Rail alongside the chat shows in real time how far the agent has gotten and how many pull requests are open.
- Choosing where to run — You can run a session on the local gateway as-is, on other hardware you've paired via
openclaw connect, or on a temporary server rented on AWS or Hetzner using the provisioning tool in the github.com/openclaw/crabbox repository. - Starting collaboration — Inviting teammates into a session lets them share the same task. According to the documentation, provider credentials always stay on the gateway and are never passed to remote machines.
| Where it runs | Characteristics | Supported backends |
|---|---|---|
| Local gateway | Default, runs immediately with no extra setup | - |
| Paired device | Runs on user-owned hardware connected via openclaw connect | - |
| Crabbox temporary cloud | Runs sessions on rented servers, discarded once finished | AWS, Hetzner |
Editor's take
The open-source AI agent field has been competing less on raw performance and more on how hard it is to even get started. Plenty of projects can already write code and generate files on your behalf, but plenty of people also gave up right at the install screen, unsure where to paste their API key. The fact that OpenClaw 2.0 poured most of its engineering effort into auto-detecting existing subscriptions and keys tells you the foundation saw that barrier to entry, not raw capability, as the real problem to solve in this release.
Anyone who's set up an open-source agent tool by hand will feel this change immediately. Copying keys into config files and losing hours to a single typo is a rite of passage for almost everyone who's used these tools. Auto-detection removes that friction, but it also means the software is scanning credentials already sitting on your machine — which is exactly the kind of thing a security team should be adding to its review checklist.
For teams considering adopting an agent tool, Shared Cloud Sessions is worth paying attention to. Most agent tools so far have assumed one person, one session. Letting multiple people share the same task context could genuinely save time for team-based adoption. That said, rather than simply trusting the design claim that credentials stay on the gateway, it's worth having someone on the team actually verify that nothing leaks to the remote machine.
Expect bug reports over the coming weeks focused on how stable Crabbox's AWS and Hetzner provisioning actually is. Renting temporary servers is convenient, but it tends to bring billing and cleanup headaches along with it — and that's likely to be the next real test for this release.
Comments