OpenClaw 2.0 Brings Guided Setup and Faster Local Control
OpenClaw 2.0, released on 30 August 2026, is a major update to the open-source personal AI assistant that runs on your own devices. It adds guided model setup that detects local Ollama and LM Studio models, a rebuilt control UI that starts in about 575 ms, SQLite storage, and a security audit command, while keeping one trust boundary per gateway.
On this page
What OpenClaw 2.0 ships
OpenClaw is an open-source personal AI assistant that runs on your own devices and meets you in the chat apps you already use, from Signal and Telegram to Discord and iMessage [2]. Version 2.0 (v2026.8.1) was announced on 30 August after an unusual pause: the project had shipped 106 releases across the previous 230 days, then went quiet for seven weeks [1].
The result bundles more than 16,000 merged pull requests from 933 contributors, roughly half the project's merge history, and it is aimed squarely at single operators and single teams rather than multi-tenant products [1].
Guided setup that finds your local models
The headline feature for local AI users is guided model setup. Instead of asking you to configure endpoints by hand, the setup pass detects what already exists on the machine: verified Codex, ChatGPT, or Claude CLI sign-ins, API keys, or local models installed through Ollama and LM Studio. It then verifies the chosen model actually works before saving anything [1].
Under the hood, the node-llama-cpp integration was replaced with a managed llama-server, Gemma 4 is now the RAM-gated llama.cpp default, and the default context window rose to 64K [1]. That lines the assistant up with what dedicated local runtimes shipped this month [1].
A faster control UI with honest security boundaries
The browser-based Control UI is now the primary surface. In the team's simulated tests with a mocked gateway at 50 ms latency, JavaScript requests dropped from 140 to 45 and startup fell from about 1.6 seconds to 575 ms [1]. The UI adds a file editor, a git-backed Changes panel, a browser panel with element inspection, and a full-screen web terminal with stated limits [1].
Security language is unusually blunt for a release note. Sessions and transcripts moved to SQLite, with a warning to back up before downgrading. Shared cloud sessions let a second person join or take over work, which the docs describe as "not tenant isolation and not a security boundary": one gateway, one trust boundary [1]. The gateway binds to loopback by default, and a new security audit command checks inbound access, tool blast radius, network exposure, and plugin allowlists [1].
What this means for local AI users
Two takeaways. First, local agent platforms are maturing fast: if you already run Ollama or LM Studio, OpenClaw 2.0 can adopt your existing models instead of forcing a parallel stack, which lowers the cost of trying a self-hosted assistant [1][3].
Second, the security posture is worth copying. Loopback by default, an audit command, and explicit statements about what a feature does not protect are exactly the honest defaults self-hosted AI needs [1]. Context from the release: a crowdsourced prompt-injection arena with 272,000 attacks scored between 0.5 and 8.5 percent attack success across popular models, while adaptive human attackers still exceeded 80 percent, so approvals and sandboxing, not model choice alone, remain the real defence [1].