OpenClaw 2.0 Ships 16,000 PRs. Your Agent Now Dreams by Default.

TL;DR

OpenClaw, the open-source personal agent that became the most-starred repository on GitHub, shipped version 2.0 this weekend. The foundation says it was built by 933 contributors (569 of them first-timers), spans more than 16,000 pull requests, and contains roughly half of every PR the project has ever merged, in one release. What started as a push to simplify setup and rebuild the browser app grew, in the announcement's own words, accidentally into a major version. The part most write-ups will skip: the release notes quietly flip several autonomy defaults to on, including model-backed background memory consolidation (the project calls it dreaming) and automatic application of self-learned skills.

The accidental major version

OpenClaw versions by calendar, so the actual tag is v2026.8.1, published to GitHub early August 31 UTC. The 2.0 branding is the foundation admitting what the diff already knew. Before this update the project had shipped, per its announcement, "106 releases in 230 days, most within a day or two of the one before them." Then came nearly seven quiet weeks, and a release that swallowed half the project's recorded history.

The scale claim holds up against the repo. Created by Peter Steinberger and now run by the OpenClaw Foundation, the project sits at 388,124 stars and 81,487 forks as of this morning, up from the 346,000 it reported in mid-August, and it has previously reported roughly half a million systems running it. The announcement, written by community manager Hannes Rudolph, says the team set out to fix onboarding and rebuild the browser app, and "doing that properly meant carrying the cleanup through the rest of OpenClaw." A cleanup that metastasizes into a major version is a familiar story; a project admitting it in the title is not.

What actually changed if you run it

The headline work is onboarding. First-time installs now start from what is already on your machine: an existing ChatGPT or Claude subscription, API keys, or local models, instead of walking you through provider bureaucracy before the first conversation. The getting-started flow is a one-line installer plus a wizard that promises a first chat in about five minutes. The browser app, where the announcement says most people meet OpenClaw, was rebuilt as a first-class client rather than a debug view with ambitions.

Beyond that, the release notes read like a small operating system changelog. Sessions can now run on paired devices or cloud workers and move their workspace with them. Shared cloud sessions add multiplayer, with named roles that scope which agents and sessions a teammate can touch. There is a searchable conversation history, conversation branching with rewind and fork, an opt-in A2A 1.0 channel for talking to peer agents, IMAP triggers so an email can start a job, experimental Docker-based fleets and parallel subagent swarms, and a Wear OS companion, because apparently the lobster needed to be on your wrist too.

The defaults that flipped

The more consequential changes are four defaults that move OpenClaw from tool that waits toward agent that acts. Per the release notes: "grounded dreaming" enables model-backed background memory consolidation by default, promoting material into long-term memory and logging it to a Dream Diary. Automatic self-learning captures reusable lessons from your sessions and applies scanner-approved skills by default. Personal installs now recall bounded context from the same agent's private conversations by default. And sessions persist across idle periods and day boundaries instead of resetting.

Dreaming is a decent name for what it does. It is the same job your brain does overnight: sorting the day's mess, deciding what mattered, filing it into long-term memory. OpenClaw's version runs a model over your accumulated context in the background and promotes what it judges worth keeping, except this sleeper keeps a diary you can read and a switch you can flip.

Every one of these ships with an explicit off switch, and the memory system gained real controls in the same release: you can inspect which sessions fed a memory, exclude sources, and delete derived memory with openclaw memory forget while keeping transcripts intact. But defaults are destiny in a project with this many casual installs. If you run OpenClaw and have never opened its memory settings, its behavior changed this weekend without you asking.

The security ledger

This is the same project whose agents keep making incident write-ups, including the Melbourne gym-booking exploit that Australian media called the country's first autonomous AI cyberattack. So it matters that 2.0's ledger is not all gas. Agents can now request credentials through a masked prompt that keeps the value out of chat and model context. Installing a plugin from an arbitrary executable source requires an explicit --force, and ClawHub surfaces available security-audit information before install. Sessions get explicit permission modes anchored to their workspace, a shared credential store keeps secret values write-only with egress bound to declared hosts, and an optional 1Password broker adds per-secret approval. Feature telemetry exists now, and it defaults to off.

The notes are also refreshingly blunt about limits, describing the new team roles as "collaboration controls, not hostile-tenant isolation." Translation: multiplayer is for people you trust, not a security boundary.

Breaking changes and a deadline

Two migrations will bite existing installs: the bundled OpenProse plugin is gone (run openclaw doctor --fix and follow the migration guide), and shipped codex/* model routes move to openai/*. Plugin authors have a harder date: a set of plugin SDK subpath imports hits its deprecation gate on September 1, 2026, which at press time is tomorrow. If you maintain an external plugin, the SDK migration guide is your evening.

The cold water

The reception is not uniformly warm. In the Hacker News thread, the recurring question is not how but why: commenters asked what they would actually delegate to it, one pointed to Google Trends interest falling off a cliff after March, and several said they had moved on to other assistants or their own builds. That skepticism coexists with the number above: 933 people sent code, 569 of them for the first time. For a project this size, the gap between the crowd building it and the crowd shrugging at it is itself the story to watch.

Key Takeaways

• OpenClaw 2.0 (tag v2026.8.1) landed August 30-31 as the project's largest release: 933 contributors, 569 first-time, 16,000+ PRs, roughly half of everything ever merged.
• The repo now sits at 388,124 stars and 81,487 forks, and versions by calendar; the 2.0 branding marks scope, not a new scheme.
• Four autonomy defaults flipped on: background memory dreaming, self-learned skill auto-apply, private conversation recall, and persistent sessions. All have off switches; check yours.
• Security got real additions: masked credential prompts, plugin provenance gates, per-secret 1Password brokering, permission modes, and opt-in-only telemetry.
• Plugin SDK deprecation gates hit September 1, and OpenProse plus codex/* routes need a doctor --fix migration.
• HN's reception was lukewarm on use cases even as contributor counts hit records; adoption and enthusiasm are diverging.

Sources: OpenClaw Foundation announcement, GitHub release v2026.8.1, openclaw/openclaw repository, OpenClaw docs, Hacker News discussion