OpenClaw, the self-hosted AI agent that grew from a weekend project by developer Peter Steinberger into one of GitHub's fastest-growing repositories, shipped version 2.0 on August 30. The release, tagged v2026.8.1, is the project's first major update aimed squarely at teams rather than solo operators, and outlets including VentureBeat, MarkTechPost and The Register covered it through September 1.

What's new

The headline feature is shared cloud sessions: an admin can now invite a colleague into a live agent session with read-only, suggest, draft or full-participation access, letting a team watch or steer a running task instead of only reviewing its output afterward. OpenClaw 2.0 also ships a rebuilt browser-based control UI that consolidates conversations, files, approval requests, configuration and live agent activity in one workspace, alongside a simplified, guided model-setup flow that the maintainers say cuts control-UI startup time to roughly 575 milliseconds.

On the security side, the release expands sandboxing options considerably: support for Docker and Podman containers, per-agent and per-session sandbox scopes, configurable read-only or read-write workspace access, role-enforced permissions, remote execution nodes, disposable cloud workers, and a team-scoped secret store that can hand a credential to a request without exposing it to the underlying model. The maintainers describe the update as touching roughly half of all pull requests merged into the project to date, reflecting the scale of the rework.

The catch: still opt-in

Coverage from The Register and other outlets flagged that OpenClaw's stronger security controls remain opt-in rather than default: the baseline configuration still assumes a single trusted operator and permits host-level command execution unless an administrator explicitly turns on sandboxing and per-request execution approvals. That mirrors a criticism that has followed the project since it went viral in late 2025 — that its default posture prioritizes convenience over containment, an especially pointed concern now that 2.0 makes it easier for multiple people to share access to a running agent.

OpenClaw remains free and MIT-licensed; the software itself has no subscription cost, though running it still requires paying for the underlying model API tokens or hardware to run local models.

Why it matters

OpenClaw's growth — reportedly north of 380,000 GitHub stars and tens of thousands of contributors and forks within roughly a year — has made it a bellwether for how far open-source, self-hosted agents can go head-to-head with commercial coding-agent products from Anthropic, OpenAI and Cursor. Version 2.0's shift toward team and enterprise workflows shows the project chasing the same "agents working together" use case incumbents are racing toward, but the persistence of insecure defaults is a reminder that broad adoption of self-hosted agents is outpacing the security tooling needed to run them safely at scale.