OpenClaw 2.0 rebuilds the personal agent for browsers and teams

OpenClaw released version 2.0 on August 30, rebuilding its browser interface, simplifying setup, and adding shared sessions for collaborative AI agent use. The open-source project, with roughly 388,000 GitHub stars and 81,000 forks, said 933 contributors, including 569 first-time contributors, worked on the update across more than 16,000 pull requests. The release follows a nearly seven-week pause after 106 releases in 230 days, aiming to make locally controlled agents more approachable while maintaining the Gateway model for credential and state control.

OpenClaw 2.0 rebuilds the personal agent for browsers and teams The open-source release simplifies setup and adds shared sessions after nearly seven weeks without an update. By Ryan Merket /author/ryan-merket · Published Primary source: OpenClaw on X https://x.com/openclaw/status/2094266903204434431 Why it matters OpenClaw is trying to make locally controlled agents approachable without giving up the Gateway model that keeps credentials and state under the operator's control. OpenClaw released version 2.0 https://openclaw.ai/blog/openclaw-2-accidentally on August 30th, rebuilding its browser interface, cutting setup work and adding shared sessions that let multiple people collaborate with an AI agent without discarding its existing context. The release pushes OpenClaw beyond the command-line enthusiasts who drove its early adoption. New installations can start with model access already present on a computer, including ChatGPT or Claude subscriptions, API keys and local models. OpenClaw has also moved optional configuration out of the initial setup flow, allowing users to finish configuring an agent through conversation. OpenClaw says 933 contributors worked on the update, including 569 first-time contributors, across more than 16,000 pull requests. Those figures are self-reported, though the scale is consistent with the activity visible in OpenClaw's public repository https://github.com/openclaw/openclaw , which had roughly 388,000 stars and 81,000 forks when the release landed. The update arrived after an unusually long pause. OpenClaw said it had previously shipped 106 releases in 230 days, typically separated by one or two days. Version 2.0 followed nearly seven weeks without a release and contains roughly half of all pull requests ever merged into OpenClaw, according to the announcement. That slowdown came from accumulated complexity rather than a lack of development. Installation, messaging, memory, skills, model support, automations, native apps, plugins and security all changed during the cycle. OpenClaw began the work as an onboarding and browser project, then expanded the scope as contributors found that those surfaces depended on the rest of the system. The browser becomes the front door OpenClaw is an open-source personal agent that connects models and tools to messaging services and devices through a local Gateway. The Gateway owns sessions, credentials and tool access, while users can interact with an agent through services including WhatsApp, Telegram, Slack, Discord, Signal and iMessage. Version 2.0 treats the browser as the primary control surface rather than a secondary dashboard. The rebuilt interface opens directly into a conversation and gives users a place to manage continuing work, configuration, files, approvals, plugins and multiple sessions. OpenClaw's getting-started documentation https://docs.openclaw.ai/start/getting-started now describes a roughly five-minute path from installation to a working browser conversation, though model authentication, channel pairing and optional plugins can extend that process. The emphasis on setup addresses OpenClaw's central distribution problem. A locally controlled agent can hold credentials, read messages and execute commands, but every additional configuration step narrows the audience to users willing to administer that access themselves. Version 2.0 attempts to preserve local control while making the first useful workflow resemble an ordinary software product. That first workflow can remain narrow. OpenClaw gives the example of monitoring a parent's inbox for school notices and sending selected updates through Telegram. The same agent can later search email for a receipt and answer an iMessage, or coordinate longer-running work across several machines. Shared sessions bring a trust boundary The larger structural change is multiplayer operation. Shared sessions allow another person to enter active work or take over a task while retaining the agent's accumulated context. OpenClaw says its own contributors used the feature while building the release. OpenClaw's cloud-session architecture https://docs.openclaw.ai/gateway/cloud-sessions separates the conversation from the machine performing the work. A session can execute on the Gateway host, a paired device or a temporary cloud worker. The Gateway remains responsible for the canonical transcript, model credentials and workspace state. Provider credentials do not need to be copied to a remote execution machine. Multiplayer support does not turn one OpenClaw installation into a general-purpose multi-tenant service. OpenClaw's security documentation https://docs.openclaw.ai/gateway/security defines one trust boundary per Gateway and recommends separate Gateways, credentials and preferably separate operating-system users or hosts for mutually untrusted participants. Shared access is designed for an individual, family or team whose members already trust one another. That limitation matters because OpenClaw agents can receive delegated authority to browse, edit files, run commands and act through connected accounts. Easier collaboration expands the number of people who can participate in a session, while the Gateway's permissions still determine the damage a mistaken instruction or compromised account can cause. Steinberger's weekend project becomes infrastructure Peter Steinberger @steipete https://x.com/steipete created OpenClaw as a weekend project after founding PDF software developer PSPDFKit. Steinberger joined OpenAI in February 2026 to work on agents while continuing to steward OpenClaw as an independent open-source effort. OpenClaw moved in July under the OpenClaw Foundation https://openclaw.ai/blog/introducing-openclaw-foundation/ , a US 501 c 3 with a full-time staff drawn from the contributor community. The foundation lists OpenAI, the University of Michigan and Offline Holdings among its major donors, while engineers from companies including OpenAI, Microsoft, Nvidia, Atlassian and Tencent participate as core maintainers. Version 2.0 is the first release large enough to show what that institutional shift buys. OpenClaw now has the contributor volume to replace major interfaces, rework its release process and introduce collaborative execution in one cycle. It also has a harder maintenance problem: an agent connected to personal accounts and real machines cannot treat upgrades like a social app shipping another feed experiment. OpenClaw's release announcement said the extra weeks were used to test both clean installations and upgrades for existing users. That is the standard the project will have to maintain as its audience moves from developers experimenting with agents to people expecting one to manage routine work without breaking the systems around it.