OpenClaw 2.0 lets multiple people run the same AI agent session

The OpenClaw Foundation shipped version 2.0 of its open-source AI agent platform on Sunday, its largest release yet at more than 16,000 pull requests, The Decoder reported. OpenClaw started as a personal project from former PSPDFKit founder Peter Steinberger and now runs on GitHub as one of the most-starred AI agent projects on the platform, with more than 388,000 stars. It works as a self-hosted gateway that connects messaging apps like WhatsApp, Discord, and Telegram to an AI agent with permission to read files, run scripts, and control a browser on the user's own machine.

• More than 16,000 pull requests merged, OpenClaw's largest release to date
• New Session Rail: in-chat status display tracking ratings, plan progress, and pull requests
• Sessions run in three places: the local gateway, the user's own paired hardware, or rented machines via Crabbox (AWS, Hetzner)
• GitHub stars: more than 388,000

The headline change is a simpler first run. OpenClaw now scans a computer for existing resources, ChatGPT or Claude subscriptions, API keys, local models, and skips the configuration steps that used to come with each one. Anything left over can get adjusted later, inside a conversation with the agent itself.

The browser app was rebuilt from the ground up. It now opens straight into a conversation and lets a user pick up ongoing work or watch a task run in real time, through a new "Session Rail" that tracks a session's ratings, plan progress, and pull requests inside the chat itself. A companion thread lets someone ask questions about what's happening without interrupting the agent's own work; per OpenClaw's documentation, that thread runs on a separate utility model reading a limited snapshot of the session.

Shared Cloud Sessions is the bigger structural change: multiple people can now work the same task together, pulling teammates into an active session or handing off a task along with its full context. OpenClaw says its own team already builds the platform this way.

A session can run in three places: the local gateway by default, the user's own hardware over a direct connection, or a disposable rented machine provisioned through OpenClaw's Crabbox tool, which supports AWS and Hetzner. Provider credentials stay on the gateway in every case and never reach the remote machine, according to the documentation.

OpenClaw's release lands inside a broader agent-tooling buildout. Intokened has tracked both sides of that trend this month: an industry-wide plugin standard for how agents plug into other software, backed by Amazon, Cursor, Microsoft, OpenAI, and Vercel, and the demand curve behind it, agent token usage on OpenRouter is now 14 times what it was in February.

The credential-isolation detail matters most for an agent with permission to run scripts and control a browser: OpenClaw is betting that keeping keys on hardware the user controls is what makes handing an AI that much access worth the risk.

Nothing here should be taken as financial advice — just information to consider.