Introduction
OpenClaw 2.0 was announced on 30 August 2026 — by a wide margin the largest update in the open-source agent's history, and one nobody set out to build. Created by Peter Steinberger, OpenClaw is a self-hosted personal AI agent that runs on your own hardware and connects to the messaging channels you already use. Shipped as version v2026.8.1, the release is aimed squarely at the two things that kept people from getting value out of it: wiring up a model, and being alone in the session once it's running.
The substantive changes are guided setup that reuses model access you already have, a rebuilt browser app, built-in memory and a complete skill-authoring path, and shared cloud sessions — the ability to bring another person into a running conversation along with everything the agent has accumulated.
The Release Nobody Planned
The announcement is titled "OpenClaw 2.0, Accidentally," and the title is the honest version of what happened. "We started by simplifying installation and rebuilding the browser app as a first-class experience," writes Hannes Rudolph, Community Manager at the OpenClaw Foundation, "but doing that properly meant carrying the cleanup through the rest of OpenClaw until it became OpenClaw 2.0."
The scale is unusual. The release was built by 933 contributors, 569 of them first-time contributors, across more than 16,000 pull requests — roughly half of every pull request ever merged into OpenClaw, in a single version. The project had shipped 106 releases in the preceding 230 days, most within a day or two of the one before; 2.0 took nearly seven weeks. The team's own diagnosis is that they were growing, and "the increased volume and pace of work outgrew both the foundation of OpenClaw and the process we used to ship it," so they reworked the foundation and the release process at the same time.
That context matters for reading the changelog. This is not a release organised around one new capability. It is a cleanup that grew until it needed a major version number, and it touches installation, messaging, memory, skills, models, automations, the browser and native apps, plugins and security.
Setup Now Finds Your Existing Model Access
Guided setup no longer starts with a blank API key field. It looks for AI access already on the machine: verified Codex, ChatGPT or Claude CLI sign-ins, an API key, a provider's own sign-in flow, or qualifying Ollama and LM Studio models. Crucially, it then proves the exact choice can answer before it keeps that model and credential — so a broken credential fails during setup rather than on your first real request.
For anyone already paying for ChatGPT or Claude, this collapses the most common failure point in self-hosted agents into a confirmation step. Configuration that used to block the first conversation was cut or moved out of initial setup, with the rest finished by talking to the agent. Local model support also expanded, adding in-process GGUF inference and llama.cpp alongside the existing Ollama path.
A Rebuilt Browser Experience
The web app was rebuilt rather than tuned, and it now opens directly into a conversation. OpenClaw reports that in a simulated default-chat test with a mocked Gateway and 50 ms HTTP/1.1 latency, JavaScript requests fell from 140 to 45 and startup fell from about 1.6 seconds to 575 milliseconds. Those are lab conditions against a mocked backend, not a promise about your hardware, but the direction is clear and the interface is also designed to stay lighter the longer it's left open — hidden panels stop fetching data they don't show.
Functionally, the interface gains full-text search across visible conversation text with the ability to reopen the surrounding messages, a durable progress card that follows subagent activity and accumulating edits across web and native chat, and widgets that can be pinned to session dashboards. The composer is built for long-running work: you can change model, reasoning or speed before sending, edit and reorder queued follow-ups, and use /btw to open a separate multi-turn side conversation without derailing the main one.
Memory and Skills Get a Real Workflow
Built-in memory now owns the core search and recall path. An eligible personal agent can recall relevant context from that agent's other private conversations — including what mattered immediately before a reset — with visible workflows to search indexed sources, inspect how recall is working, import history, and remove attributable derived memory. Recall stays inside the same agent's private conversations and respects explicit isolation policy. Users of the previous QMD backend need openclaw doctor --fix; QMD-only reranking, query expansion and cross-agent transcript search are retired.
Skills — reusable instructions the agent can follow again — get the whole path connected: create, validate, install, call directly from a conversation, review proposed changes, and have supported edits live on the next turn of a persistent session. A new Skill Workshop collects proposals, checks and applied history in one place, and self-learning can turn substantial work and durable corrections into proposed skill improvements where enabled. Invalid skills are now reported individually instead of taking the catalogue down with them.
Shared Sessions Are Cloud-Backed
Shared sessions are the genuinely new capability, and the one with the most caveats. They exist because the team hit the problem internally: "OpenClaw had no way to bring another team member into the work without losing what the Claw already knew."
On a multi-user Gateway, a conversation keeps its creator and each identified person's prompts visible, and owners or administrators choose whether someone else may read, suggest changes, work in a draft, or participate directly. The point is handing over the task with its context intact — the agent's memory of what it tried and changed, rather than a pasted summary.
Two things are worth being clear about. First, sharing runs through OpenClaw's cloud; the rest of the system is still local-first, but this feature is not — which sits awkwardly beside the announcement's own pitch that it is "not selling anything here or asking you to trust one company, one model, or one AI provider." Second, the release notes say plainly that these controls "are not tenant isolation or a security boundary, and revoked access can briefly look available until the UI refreshes or the Gateway rejects the action." Treat shared sessions as collaboration among people who already trust each other, not as access control.
Storage, Privacy and Breaking Changes
Sessions and transcripts move from files into SQLite. That is close to a one-way door: before downgrading to an older file-backed release you must use the current CLI to restore archived legacy transcript artifacts, and sessions created after the migration will not appear in older releases. Create a verified backup before upgrading.
Security work includes a team-scoped Secret Store that separates protected values from agent-readable ones, so a credential can be requested through a masked prompt, referenced from Vault or 1Password, and substituted into an approved Gateway-hosted HTTPS request without ever entering plaintext configuration or model-visible text. A new incognito mode is, in the release notes' own words, "off by default and deliberately narrower": the conversation stays in process memory and skips the normal transcript, but the model provider still receives every message and tools can still write files and reach external services.
Other breaking changes: the OpenProse plugin is removed, OpenAI model routes migrate from codex/* to openai/*, SDK import paths changed, and fresh CLI installs move to Node 22.22.2.
Conclusion
OpenClaw 2.0 is a distribution release more than a capability release. Nothing here changes what the underlying models can do; what changes is how many people can get an agent running, how much it remembers between conversations, and whether a session has to stay with the person who started it. Credential auto-detection removes the step where most self-hosted agent installs die, and shared sessions turn a single-player tool into something closer to multi-agent teamwork with humans in the loop.
The caution belongs on the sharing feature. It routes through the cloud, its permissions are explicitly not a security boundary, and the storage migration is hard to reverse. Back up first, and share sessions only with people you would already trust with access to the machine.