OpenClaw 2.0 just dropped (what changed?)
What's actually new in 2.0
OpenClaw 2.0 dropped three hours ago and the whole point of this release is control. The browser app got rebuilt, so it now has conversation search, persistent progress cards, an editable prompt queue, and chat branching. Sessions can run beyond your gateway on a paired device or temporary cloud worker while the gateway keeps the canonical transcript, model credentials, and core state. Context moved from flat files to SQLite, with active memory and background consolidation. Collaboration is now a first-class feature: shared sessions with view-only, suggest, draft, or direct participation modes, plus scoped approvals for automations.
The everyday workflow change matters most. Before, agent runs felt fragile: you sent a task, it ran somewhere, and the status was unclear if you interrupted or handed off. Now you get a durable progress card, an editable queue, and searchable branching to recover from mistakes. You can actually see what's going on.
Security: better controls, same architecture
The reason we left OpenClaw for Hermes agent was security. The gateway could touch shell tools, browsers, files, credentials, plugins, and messaging channels all at once. Plugins are executable code, and shared setups made access unclear. In 2.0, permissions and provenance are more explicit. Credential requests are now masked, so API keys don't appear in chat or model context. There's a team-scoped secret store in SQLite, with optional 1Password broker integration and value-free audit records. Plugin installs surface capability, source, version, and artifact, and risky sources require an explicit force flag.
Remote workers now execute tasks without dumping API keys onto ephemeral machines, idle workers suspend, and the gateway keeps credentials centrally. Role-based access and sandbox execution options exist. But the root architecture hasn't changed: OpenClaw is still gateway-first, and any tool that reads a secret directly can expose it to the model during that run. The improvements are real, but they're mostly at the visibility and approval layer rather than a fundamental redesign.
OpenClaw 2.0 vs. Hermes agent
For our setup, we're not moving back. Hermes agent is more secure and its ecosystem is agent-first, while OpenClaw is gateway-first. OpenClaw's channel connectivity advantage has narrowed as Hermes improved, and we've started using Lark as an all-in-one communication hub for our agents. The OpenClaw team made meaningful progress, but comments on the release show plenty of users still aren't satisfied. If you need durable context, branchable sessions, and remote worker orchestration, OpenClaw 2.0 is worth a serious look. If security posture and agent-native design are your priority, Hermes remains the better daily driver.