OpenClaw 2.0 has arrived after a seven-week pause, bringing a major reset to the project after 106 releases across the previous 230 days. The OpenClaw team used that break to rewrite installation, rebuild the browser Control UI, move sessions and transcripts into SQLite, and add shared cloud sessions.
The result is more than another release in a fast-moving sequence. OpenClaw 2.0 changes how people enter the system, how they work with conversations, and how multiple people can join live agent work while keeping the surrounding trust boundaries visible.
Setup Now Meets Users Where They Already Work
OpenClaw 2.0 turns setup into a guided path rather than a single configuration choice. The process can reuse verified Codex, ChatGPT, or Claude CLI sign-ins, accept an API key, run a provider sign-in, or find installed Ollama and LM Studio models.
That range matters because users can bring existing subscriptions, credentials, or local models into the setup flow. Guided setup also verifies the model before saving it, giving users a check before the configuration becomes part of their OpenClaw environment.
Fresh OpenAI setups now default to GPT-5.6. On the local-model side, OpenClaw replaced node-llama-cpp with a managed llama-server, made Gemma 4 the RAM-gated llama.cpp default, and raised the llama.cpp default context to 64K.
These changes connect model selection to the first moments of installation. Instead of treating setup as a separate technical hurdle, OpenClaw 2.0 brings provider sign-ins, API keys, local model discovery, and model verification into the same guided experience.
A Faster Control UI Puts Conversations First
The rebuilt Control UI places conversations at the center of the workspace, with files, approvals, and live work beside the chat. That layout makes the conversation the starting point while keeping the surrounding actions visible in the same interface.
The performance numbers show how much the rebuild changed under the test conditions. In a simulated default-chat test with a mocked Gateway and 50 ms HTTP/1.1 latency, JavaScript requests fell from 140 to 45, while startup dropped from roughly 1.6 seconds to 575 ms.
That is a major reduction in both request traffic and wait time. The test-harness startup now reaches 575 ms, giving the Control UI a much quicker path from launch to an active chat workspace.
Storage also moved to SQLite, bringing sessions and transcripts into the new storage system. The migration has an important rollback rule: before returning to an older file-backed release, the current CLI must restore archived legacy transcript artifacts.
Sessions created after the migration will not appear in older releases. That detail makes the storage change part of the operational story, not just an internal implementation update, because moving between release generations requires care.
Shared Sessions Add Multiplayer Without Replacing Security
Shared cloud sessions let a second person join live work or take it over with context intact. Owners and administrators choose what that person can do: read, suggest changes, work in a draft, or participate directly.
That permission model gives shared work a set of defined participation levels. A person can observe, offer input, prepare changes, or take part directly, depending on the access selected by the owner or administrator.
But shared cloud sessions are not a security boundary. The feature adds real multiplayer collaboration, yet access decisions still depend on the surrounding controls and the permissions granted to each participant.
OpenClaw keeps the Gateway bound to loopback by default, and most chat channels answer an unknown direct-message sender with a pairing code. OpenClaw security audit checks inbound access, tool blast radius, network exposure, browser control exposure, and plugin allowlists.
Model choice is treated as the first mitigation for prompt injection. OpenClaw cites a 2026 crowdsourced arena covering 272K attacks across 41 agent scenarios, with scores counted only when an agent both executed the harmful action and hid it from the user.
- Claude Opus 4.5 recorded 0.5% success.
- Sonnet 4.5 recorded 1.0% success.
- Haiku 4.5 recorded 1.3% success.
- Gemini 2.5 Pro recorded 8.5% success.
Those figures still sit inside a tougher warning: adaptive human attackers exceed 80% success against state-of-the-art defenses. That contrast places the security work beside the model choice, Gateway defaults, pairing flow, browser controls, and plugin allowlists rather than treating any one feature as a complete answer.
A Large Release Built Around Practical Control
OpenClaw 2.0 arrives with a clear focus on control: control over setup, model selection, conversation speed, stored transcripts, shared access, and security exposure. The project also points to Key Takeaways v2026.8.1, which lands 16,000+ PRs from 933 contributors.
The release pause now looks like preparation for a broad rebuild. OpenClaw moved from a rapid run of 106 releases in 230 days to a version that reshapes its installation path, browser workspace, storage layer, collaboration model, and security checks.
The next phase will test how these pieces work together in real use. A 575 ms Control UI startup, guided model verification, SQLite sessions, and live shared work give OpenClaw 2.0 a stronger foundation, while the security boundaries make clear that collaboration and automation still require deliberate access decisions.
