OpenClaw shipped version 2.0 over the weekend. Its lead change is a shared cloud session: a colleague can step into an agent’s work mid-task, with the context built up so far intact.
The project is open-source software users host themselves, turning a large language model into an agent that carries on after the chat window is closed, wakes itself on a schedule of its own, and can be reached through Signal, Discord, Telegram or WhatsApp.
The release is plainly aimed at teams and enterprise buyers. Peter Steinberger, who built OpenClaw before joining OpenAI in February to lead its personal-agent work, wrote on X that his own team had swapped per-developer local coding harnesses for one shared setup at team.openclaw.ai, calling local harnesses “relics of the past.”
Setup changed as well. It now opens by looking for credentials the user already holds — an API key, a Claude or ChatGPT subscription, an Ollama model running on the machine — and only then proposes setting up something new. It also checks that the configured model actually works before storing it, and drops the user into a real conversation rather than a settings page.
That conversation is where the software now lives: chat holds the centre of the rebuilt Control UI, in place of a standalone overview dashboard, giving it an appearance close to ChatGPT, Claude and Gemini. A terminal, a browser pane, pull request status, Git diffs and files are ranged around it, sparing users the move between separate tools.
The release is outsized by the project’s own standards: 933 contributors, 569 of them taking part for the first time, merged more than 16,000 pull requests into it — about half of everything OpenClaw has merged across its entire history. None of it was planned: the team set out to simplify installation and rebuild the browser app, and its own announcement calls what followed accidental.
Existing users upgrade by the usual command-line route, and the release notes warn them to back up sessions before the shift to new SQLite-based storage: a session created after the move cannot be recovered if they later revert without a backup.
Security grew alongside the multi-user features: approvals granted per individual request, command permissions narrowed to particular arguments, sandboxing through Docker and Podman, execution governed by roles, and a team-scoped Secret Store that can hand a credential to a request while keeping it hidden from the model. Two of those, the sandboxing and the per-request approvals, ship turned off.
In its own documentation, OpenClaw says a single Gateway amounts to a single trust domain, not a wall between users who lack trust in one another. True tenant isolation means standing up separate Gateways — cells, in OpenClaw’s terminology — rather than flipping a setting.