OpenClaw 2.0: Biggest update yet makes the AI agent easier to install

What this is about

OpenClaw, the open-source, self-hosted AI agent, has reached version 2.0 (officially version 2026.8.1). The project calls it "by far the biggest update in OpenClaw's history." OpenClaw was formerly known as Clawdbot and later Moltbot. It went viral at the end of last year and is now managed by a nonprofit foundation backed by OpenAI. According to Mashable, founder Peter Steinberger was hired by OpenAI.

The foundation says the update was entirely crowdsourced: 933 contributors, 569 of them first-timers. It consists of more than 16,000 pull requests (proposed changes to the code). heise reports nearly 17,000 and says that amounted to roughly half of all changes accepted in the project up to that point.

Easier installation and a new browser app

Mashable reports that the project admits the original OpenClaw was more complicated for beginners than it needed to be. So settings were moved out of the installation, simplified or removed entirely.

On fresh installs, OpenClaw now detects existing access on its own. According to heise and the Times of India, that includes existing ChatGPT or Claude subscriptions, API keys, Codex logins and local models, for example via LM Studio and Ollama. You configure anything else later by chatting with the agent. Setup has been simplified on the desktop (Linux, Windows, macOS). The iOS and Android apps get a leaner flow for permissions and device pairing.

The browser app was rebuilt from scratch. It opens straight into a conversation. There you continue setup, pick up running tasks and watch the agent work live. Approvals, files, settings and chats sit in a single view. You also get pinned chats, grouped sessions, cross-session search and real-time Markdown rendering. Live cards show task progress, and you can search past conversations.

Team features, memory and computer control

The project calls Shared Cloud Sessions "multiplayer." Several people can join a running session or hand off tasks without losing the existing context. According to heise, the team now uses the feature itself in development.

The detailed release notes include more, according to heise:

• Active Memory: On personal installs, the agent can retrieve limited information from earlier private conversations.
• Grounded Dreaming: A default feature meant to condense relevant content into long-term memories in the background.
• Self-learning system: It can record reusable insights and automatically apply vetted skills, including ones from the Skill Workshop.
• Automation: Tasks stay linked to the conversation in which they were created. For recurring actions, a one-time approval is enough from now on. Incoming emails can trigger agent work through an IMAP plug-in.
• Computer control: On connected Macs, the agent can operate desktop interfaces. On Windows and Linux this is experimental and must be explicitly enabled.
• Conversation history: You can search, branch or rewind conversations. File changes or tool actions that have already run are not undone in the process.

According to the Times of India, you can set up routines that watch incoming mail, flag urgent messages and send push notifications via Telegram. The agent can also work across channels, for example finding an invoice in your inbox and replying with the details via iMessage.

Security: progress, but core risks remain

There is a new feature for private credentials: the agent can request them without they appearing in plain text in the chat or the model context. Access rights for shared sessions can be set more precisely, and there are new protections for plug-ins. According to InfoWorld, the update also changed runtime behavior, isolation, the plug-in lifecycle and the controls over how agents interact with external tools.

heise notes that the blog post mentions security only in passing, even though it remains a core issue. OpenClaw itself names prompt injection as a danger: an agent processes untrusted content while also having access to tools with far-reaching permissions. Sandboxing is off by default. Mashable also warns that you should understand the risks before handing your device over to the agent.

Jaishiv Prakash, Director Analyst at Gartner, told InfoWorld: "The latest version of OpenClaw shows that agent security can't be solved separately for runtime, memory and integrations." These parts share identity, context and permissions, so one weakness can compromise the whole workflow. Companies need to check whether execution, memory, credentials and connectors are protected by enforceable trust boundaries and whether each agent gets only minimal privileges.

Background: why nothing shipped for so long

According to the project, there were 106 releases in 230 days before this update, usually one to two days apart. The nearly seven-week pause was therefore unusual. The team had grown, and the volume and pace of work overwhelmed OpenClaw's foundation and its release process, so both were reworked at the same time. What began as a simplification of the installation and browser app turned into an overhaul of almost every part: messaging, memory, skills, models, automations, native apps, plug-ins and security. The team joked on social media: "HA! We beat GTA 6!" The Times of India headline claiming Sam Altman spent "billions" isn't backed up by the other sources.

What you need to try it

OpenClaw is an agent, not a model, so it doesn't need a GPU of its own. There is an installer for Mac and Windows on the OpenClaw website. Developers can set it up themselves. The agent uses an existing ChatGPT or Claude subscription, an API key, or a local model via Ollama or LM Studio. Only local models require plenty of VRAM (video memory) or RAM. The sources give no precise hardware requirements. It's best to test it on a separate machine or in a sandbox.

Our take

OpenClaw 2.0 is above all an accessibility update. Mashable notes that the project lost ground because other vendors offer similar agent features that are easier to use. According to heise, its strengths compared with OpenAI, Google, Anthropic and Microsoft lie in open source, self-hosting and free choice of model. Microsoft's Agent Scout is built on OpenClaw, according to heise, and Nvidia supplies additional security and sandbox mechanisms with NemoClaw. That makes the project a building block in the agent ecosystem.

The critical points remain: security still depends heavily on configuration, sandboxing is off by default, and computer control on Windows and Linux is still experimental. How well the new protections and the overhaul work in everyday use will have to be shown by independent tests, which the sources don't yet include. The figures on contributors and pull requests come from the project itself and differ slightly between sources.