What changed in OpenClaw 2.0, released on August 30 as the project’s largest update after a nearly seven week, two month rebuild involving 933 contributors—including 569 first time contributors—and more than 16,000 pull requests, and how does it improve installation, model and su

OpenClaw 2.0 (v2026.8.1) is a broad platform rebuild rather than a single feature release: it modernizes setup, persistent state, the web control plane, team workflows, and security boundaries. The project describes it as its largest release, built from more than 16,000 pull requests by 933 contributors, 569 of them first-time contributors. 3

What changed

• Installation and model setup: the installer is redesigned to guide configuration, discover available model credentials and existing subscriptions, and reduce manual provider setup. This is intended to make a working first agent easier to provision across supported providers. 3

• Messaging, memory, and sessions: conversations, session history, and transcripts move from flat JSON/JSONL-style storage toward SQLite-backed persistence. That gives OpenClaw a queryable, more durable basis for conversation search, memory, archives, session handoff, and shared state. Earlier documentation identifies the former sessions.json store and *.jsonl transcripts. 7

• Skills, models, and automations: the release expands reusable/self-improving skills, model selection and provider handling, and recurring or persistent automated work. It also puts the Gateway at the center of session routing and channel connections. 8

• Browser Control dashboard: Control is substantially rebuilt, with faster startup—reported as about 575 ms rather than roughly 1.6 seconds—and richer operational tooling. The UI includes configuration forms generated from the live schema, including plugin and channel schemas, while retaining a raw-JSON escape hatch. 6

• Developer/operator panels: the redesigned dashboard adds workspace-oriented controls such as file editing, Git/PR/CI visibility, browser inspection, and web-terminal workflows, enabling more work to be inspected and controlled without leaving the browser. 3

• Native apps and plugins: native-client and paired-device workflows are strengthened, while plugin configuration becomes more visible and manageable through the Control UI’s schema-driven settings experience. 6

Teams and shared cloud sessions

• Shared sessions let work run on paired devices or cloud workers, carry the workspace with the session, reuse warm machines and project seeds, and let teammates join or take over a task with its existing history and context rather than restarting it. 23

• This is a collaboration and continuity feature, not a security boundary. Teams should not treat shared sessions as tenant isolation or as a substitute for separate gateways, credentials, access controls, or infrastructure boundaries. The project’s architecture still describes the Gateway as the source of truth for routing and channel connections, with isolation configurable by agent, workspace, or sender. 8

• The release adds a shared, team-scoped credential store in SQLite. Secret values are write-only, protected outbound access is limited to declared hosts, and readable environment values are distinguished from secrets. 3

Security and upgrade caution

• The upgrade is significant enough that users should create and verify a backup of configuration and state before updating. OpenClaw’s backup tooling supports archive creation, verification, restoration, and SQLite-specific backup operations; backups may include state, configuration, authentication profiles, provider/channel credentials, sessions, and optionally workspaces. 13

• This matters especially because the persistence migration changes where and how sessions and transcripts are stored. Do not rely on copying only legacy flat files as a migration strategy; use the supported backup and restore path. 17

What remains true

• The intended direction remains an open, portable agent platform: users can choose models, providers, devices, plugins, and deployment environments rather than being locked to one AI vendor or company. The available evidence supports the technical changes above, but is insufficient on its own to independently verify the exact MIT-license and governance wording.