The open-source lobster moves to version 2.0
OpenClaw 2.0 simplifies installation, rebuilds its browser interface, and adds improved memory, automations, shared sessions, and new safeguards.
Nearly two months of development turned an installation cleanup into a complete overhaul. OpenClaw has published its largest update under version number v2026.8.1, also known as OpenClaw 2.0. According to the foundation, the release brings together more than 16,000 pull requests from 933 contributors, including 569 people contributing to the project for the first time.
Those figures reflect the project’s unusual release pace. OpenClaw had previously shipped 106 versions in 230 days, often only a few days apart. The pause before this release allowed the team to rework installation, the browser interface, memory, automations, plugins, native apps, security, and session management at the same time.
The first change affects how users enter the service. During a new installation, OpenClaw detects more resources already available on the computer, including compatible ChatGPT or Claude sign-ins, API keys, and locally hosted models. Some settings have been removed from the initial setup or moved until after the first conversation. Users can then continue configuring their assistant by speaking directly with it.
Support for existing subscriptions does not mean that a ChatGPT or Claude plan becomes universal access to the corresponding APIs. OpenClaw separates subscription-based authentication from usage billed through an API key. For OpenAI, a ChatGPT/Codex connection uses the quotas and credits associated with the account, while an OpenAI Platform key remains separately billed. Available methods also depend on the rules established by each provider.
The browser interface becomes the main entry point. It opens directly into a conversation and allows users to complete setup, return to previous sessions, follow ongoing work, inspect related files, or respond to approval requests from the assistant. A search function can also retrieve older conversations through exact words or phrases.
This browser app is not a conventional hosted service. It is served by the OpenClaw Gateway installed on the user’s computer or server. Locally, it runs by default through an address limited to the machine. Remote access requires additional configuration, including authentication, approval for new devices, and a recommended HTTPS connection.
Task monitoring becomes easier to follow. Persistent cards remain available after the page reloads and display session progress, completed actions, accumulated changes, and activity from secondary agents. Interactive results can also be pinned to dashboards, with specific permissions controlling the available actions and the domains they can access.
Memory receives several changes. With Active Memory enabled, a personal installation can retrieve selected information from previous private conversations with the same agent. Group conversations and shared channels remain excluded from this behavior by default. A background consolidation system can also move information with traceable sources into long-term memory. Users can restrict or disable these features.
OpenClaw 2.0 also introduces automatic learning from usage. Lessons considered reusable can be converted into skills and applied after verification. Changes written directly by the user remain pending until approved, while different control levels allow this behavior to be disabled or limited to suggestions.
Automations can now remain attached to the conversation where they were created. Recurring authorization can be granted for one precisely defined operation, then reviewed or revoked later. Any change to the task or authorized operation requires new approval. OpenClaw also adds masked credential requests, keeping sensitive values out of the conversation and away from the model’s context.
A session can move from the main computer to a paired device or remote infrastructure while retaining its workspace. Machines that have already been prepared can be reused, avoiding a complete setup process whenever another task begins.
Shared cloud sessions extend this continuity to collaborative work. Another person can join an active task or take it over without losing the context accumulated so far. Defined roles determine which agents, sessions, and operations each person can access. The documentation notes that these controls are intended for teams whose members trust one another, rather than for fully isolating potentially hostile users.
This collaborative layer includes a shared area for credentials and environment variables. Secrets remain unreadable after being stored, and their use can be restricted to approved destinations. An optional 1Password connection can also retrieve selected information after authorization. In certain workflows, however, a requested value may still be exposed to the model during that specific execution, which should be considered before adding critical credentials.
Plugin management becomes more explicit. Before installing or updating an external plugin, the interface can display its source, version, requested capabilities, and the exact artifact involved. Arbitrary executable sources require an additional confirmation, while available ClawHub audit information can be reviewed before installation.
The release also introduces several breaking changes for existing configurations. The bundled OpenProse plugin and `/prose` command have been removed in favor of migration to a separate skill. Previous `codex/` and `openai-codex/` references must move to the new `openai/` structure. Several interfaces used by external plugins are also changing paths.
OpenClaw recommends backing up the Gateway configuration and state before upgrading. The `openclaw doctor --fix` command can repair some legacy references and remove outdated settings, but certain conflicts may still require manual review. An existing installation therefore requires more care than a fresh setup.
The project remains distributed under the MIT License. Its code can be inspected, modified, and redistributed, while users retain the option to run the Gateway on their own