OpenClaw 2.0 Is Here: What's New, What Breaks, and What It Signals

§ 01The release in numbers#

The announcement’s own framing is worth keeping: this was not a planned major version. The project’s velocity outgrew its foundation and its release process at the same time, and the team reworked both rather than shipping around the problem.

§ 02What’s actually new#

Setup that starts with what you already have. Guided onboarding now detects AI access already on your machine: active ChatGPT or Claude subscriptions, API keys, and qualifying local Ollama or LM Studio models. It verifies that the exact chosen model can answer before saving it, then hands off to the rebuilt browser app, which opens directly into a conversation. Most configuration moved out of initial setup entirely; you finish configuring your Claw by talking to it.

Multiplayer, for real. Shared cloud sessions are the headline: team members can join an ongoing agent session, continue work where it left off, assign tasks to teammates with full context, and collaborate live. The OpenClaw team says it now uses this multiplayer mode to build OpenClaw itself. New team operator roles limit which agents and sessions each verified person can reach, and the docs are refreshingly honest about scope: these are collaboration controls, “not hostile-tenant isolation.”

Memory grows up. Built-in Memory now owns the core search and recall path, with the QMD add-on retired via a supported migration. Two defaults are worth noticing: background memory consolidation promotes provenance-qualified material into long-term memory (with a Dream Diary and an explicit off switch), and automatic self-learning can capture strong reusable lessons and turn them into proposed skills. There is also plain conversation search at last: find past conversations by exact words and reopen the surrounding messages.

Skills become a workflow. Skill Workshop connects the whole path: create a skill down a guided route, validate it before anything is written, find or install skills, call them from a conversation, and review proposed changes with an applied history. Invalid skills are now reported individually instead of breaking the catalog around them.

Widgets and dashboards. Chat can now hold interactive widgets, pin them to session dashboards, grant them specific actions or network origins, and export rendered views as images. Combined with durable session progress cards, the browser app is now a real control surface rather than a log viewer.

A more structural security model. This is the deepest change for anyone running OpenClaw seriously. Approvals now stay attached to the exact request, command, session, and person that received them. Reusable command permissions bind to exact arguments and a working directory, and script-backed commands recheck the reviewed bytes before running. Each session can choose read-only, guarded, workspace, or full access. A shared credential store keeps team-scoped secret values write-only, agents can request credentials through masked prompts that never enter model-visible text, and plugin installs from arbitrary executable sources now require an explicit force flag. After July’s advisory batch, this is the follow-through release: the trust boundaries the project promised are now load-bearing features rather than policy statements.

§ 03What breaks#

Three migrations, all documented, plus one date to diary:

The pattern across all four is the same: openclaw doctor --fix is the recovery tool, and the release notes flag conflicts for operator repair instead of silently guessing.

§ 04What the release signals#

OpenClaw 2.0 is a strong release, and the honest read is that it validates the direction this whole category is moving. A year ago, “personal agent” meant a chat loop with tools. The features that earned 2.0 status are all continuity features: memory that consolidates in the background, skills that accumulate from real work, permissions that persist across sessions, and shared context that survives a handoff between people. Those are the properties of a colleague, not a chatbot - the distinction we mapped in AI agent vs AI employee.

The difference is who operates it. OpenClaw is open source and self-hosted: you own the Gateway, the credential store, the upgrade path, and the security perimeter, and 2.0 makes each of those meaningfully better while leaving them yours to run. The announcement is admirably direct that this is the point: no company to trust, software you can shape and actually own. For tinkerers and teams with an operator’s mindset, that is the draw, and 2.0 lowers the cost of entry more than any release before it.

CellCog sits on the other side of the same conviction, so read our position with that declared: we build AI employees where the platform owns the perimeter, the memory architecture, and the upgrade path, and the two worlds already cooperate - OpenClaw agents have used CellCog as an execution layer for heavy deliverables since early 2026, and that path is unchanged in 2.0.

§ 05If you run OpenClaw today#

The upgrade checklist, condensed from the release notes:

• Create a verified backup before upgrading; review the SQLite downgrade notes if you might roll back.
• Run openclaw doctor --fix after upgrading to migrate OpenProse, QMD memory settings, and legacy codex/* routes.
• Review the new per-session permission modes and set the narrowest that works; the least-privilege playbook applies directly.
• Check any external plugins against the September 1 SDK deprecation gates.
• Re-run your hardening pass: the July advisory checklist still applies, and 2.0’s new controls give it sharper teeth.

A release this size will shed follow-up fixes for weeks. When the notable ones land, this page gets the update, same day.

§ 06Update log#

This is a living page; follow-up fixes and advisories land here.

August 31, 2026 - the first day-one regressions, all upgrade-path bugs. Within 24 hours of release, four notable issues were filed on the OpenClaw tracker, and the pattern is consistent: the new features hold up, the migration tooling is where it creaks.

• doctor --fix silently skips 2.0 migrations without a TTY (issue 134036, P1, filed and already closed with a fix). Run over ssh or from automation with no terminal attached, doctor --fix and --repair skipped the doctor-owned state migrations entirely, while every error message kept telling operators to run exactly those commands. Reported fallout included Gateway crash loops and blocked exec approvals. Until the fix ships in a patch release, the workaround is to run repair through a real or pseudo-TTY.
• Memory sync aborts on Gemini embeddings (issue 134015, P1, open). The embedding path batches more than 100 requests per call, Google rejects the request, and the sync aborts instead of splitting the batch, which can leave agent memory stale.
• Plugin consent not persisted for legacy installs (issue 134076, P2, open). plugins update --all --accept-capabilities can report success while leaving older plugins blocked pending per-plugin re-consent.
• Dashboard readiness false positive (issue 133953, P2, maintainer-filed, closed). A raw WebSocket close could be read as a reachable Gateway, offering to start a second service on a busy port.

As of this update, no patch release has been published; a maintainer comment on the release thread points to one within roughly a day. The takeaway for operators matches our checklist above, with one sharpened line: run your upgrade from an interactive terminal, not from a script, until the TTY fix lands. Filed dates and statuses verified against the GitHub tracker on August 31.

September 1, 2026 - the patch release, v2026.8.2. Published September 1 at 16:00 UTC on GitHub (npm at 16:19 UTC), about 33 hours after 2.0. By the team’s count: 784 pull requests, 10 direct commits, 134 contributors. Here is where the four day-one issues stand against it, verified on the tracker and in the release notes on September 2:

The release’s own headline for operators is the “Safer upgrades” line: preserve newer configuration, stop incomplete session migrations before claiming success, and recover a stopped Gateway after a failed update when the installed package or rollback is verified safe. That is the migration-tooling class the day-one reports pointed at, addressed structurally rather than issue by issue. Also in 8.2: a Home agent that opens in a dock beside your current page, a Linux desktop companion, background sessions started without switching pages, browser control that works without a running Gateway, and four new Control UI themes.

Two lines matter for anyone who read our security post. First, a default changed: unsandboxed sessions can now work with other sessions of the same agent by default, including retained cron sessions; the notes tell shared-agent operators to set tools.sessions.visibility to tree or self when they need narrower access, and sandbox and cross-agent restrictions stay enforced. Second, the managed Sharp dependency moved to 0.35.4 with libheif 1.23.2 to fix image-decoding vulnerabilities, with one known exposure left open: standalone @openclaw/memory-lancedb installs can still resolve an older Sharp through an optional dependency, which the team lists as a packaging follow-up. Plugin authors should also note that the deprecated Plugin SDK import paths remain available in 8.2 with a recorded removal target of September 1, per the notes, so the migration guide is now overdue reading. Verified against the GitHub release, the release notes at docs.openclaw.ai, and the tracker on September 2.

September 3, 2026 - the third release in four days, v2026.9.1. Published September 3 at 18:31 UTC on GitHub, two days after the 8.2 patch. By the team’s count: 1,186 pull requests, 28 direct commits, 281 contributors. This is not another patch; it is a full monthly-numbered release, and its center of gravity is the thing the day-one reports exposed: upgrades that leave you working.

What matters for operators, from the release notes:

• openclaw update now rolls back the npm candidate when the post-update Doctor fails, preserves configuration and secret references across a failed upgrade, hands failures to a built-in triage agent, and waits for plugin readiness before restarting. One instruction for anyone who upgraded to 8.2 on a machine without a service manager: run openclaw update --no-restart once.
• The Gateway stays up through bad state: malformed legacy cron rows are quarantined instead of blocking boot, migration warnings degrade the Gateway rather than refusing to start, and Windows Gateways survive an agent restart. Two new knobs land with it: cron.skipMissedJobs skips missed recurring jobs at startup, and blockedHostnames on the SSRF policy blocks configured hosts for browser, web fetch and webhooks.
• Approvals get durable and visible: “Allow Always” sticks for MCP tools on OpenClaw-configured servers, and when a delegated system agent proposes a config change or restart, the approval card reaches the channel that asked.
• Setup and surfaces: a one-prompt quick-start lane for fresh installs that detects existing Claude Code or Codex logins and API keys; Mermaid diagrams rendered in every chat across Control UI, macOS, iOS and Android; personal skill libraries on shared Gateways; Android brought up to the web UI. Models: Fable 5.1 support from shared metadata, SuperGrok usage shown, Codex managed installs moved to 0.152.1.
• Leaner installs: unused Sharp and ONNX dependency chains dropped, with Chromium or Chrome now required for meme PNG output.

Against the day-one list: the plugin-consent persistence issue (134076) is referenced in the plugin fixes; the Gemini-embeddings memory-sync abort (134015) is not named in the notes, so treat its batch-size workaround as still standing. The Plugin SDK deprecation removal is also not named, which means the deprecated import paths appear to survive another release past their September 1 target; check your plugins before assuming either way. Verified against the GitHub release and the release notes at docs.openclaw.ai on September 3.

September 5, 2026 - the fourth release in six days, v2026.9.2. Published September 5 at 20:00 UTC on GitHub, two days after 9.1. By the release’s own audited record: 1,247 merged pull requests since 9.1; the notes credit contributors by name but publish no total this time. The release-notes index at docs.openclaw.ai sums it up as “Reliability and recovery improvements, OpenAI GPT-6 Astra and Meta Muse Spark 1.3 support, and flexible task workspaces.” Three things stand out.

• GPT-6 Astra is a model option. openai/gpt-6-astra works with an OpenAI API-key profile or an eligible ChatGPT or Codex account, with text and image input, Responses tool calls, reasoning controls and a new /think ultra level; on Platform API-key routes it also runs function tools asynchronously and steers active responses mid-turn. Subscription access “follows successful account discovery”, so whether your ChatGPT plan reaches it depends on OpenAI’s rollout, which our Astra tracker follows. That makes OpenClaw an Astra harness two days after launch; Cursor, by OpenAI’s own August 28 announcement, will not be one.
• Two defaults widened, and both matter for anyone who read our security post. Swarm is enabled by default: concurrent sub-agents with structured results and live progress, with the existing opt-outs, tool restrictions and the separate Code Mode opt-in preserved. And session tools now default to all-session visibility with ordinary agent-to-agent access enabled; 8.2 had opened same-agent sessions, 9.2 opens the rest. The notes name the narrower settings, tools.sessions.visibility set to agent or self, and say tool and sandbox restrictions still hold. Shared-Gateway operators should set that before upgrading, not after.
• The upgrade story continues. Automatic updates keep active settings, enabled skills and default-agent ownership; Gateway restarts are restored after Git updates; active, queued and delegated replies survive a Gateway restart; corrupt backup archives are rejected instead of accepted incomplete; and more settings apply without a restart, with the ones that still need one marked in the configuration reference.

Also in 9.2: faster chat and dashboards while long transcripts are processed, macOS browser sign-in through Cloudflare Access, an experimental plugin UI (Settings, Labs), standalone Talk on Apple Watch (experimental), personal connected accounts, dashboards browsed as a gallery, Discord meeting notes, and Owner profiles for single-user Gateways.

Against the day-one list: the Gemini-embeddings memory-sync abort (134015) is still open on the tracker, so the batch-size workaround stands for a fourth release; the plugin-consent issue (134076) was closed September 1. One deprecation now carries a date: the Plugin SDK’s untrusted-named prompt-context aliases become eligible for removal on or after September 8, 2026, so migrate to the channel-named context fields before then. The release verification block says ClawHub is still recovering failed staging records and the signed macOS assets await approval; if openclaw update cannot see 9.2 yet, that is why. Verified against the GitHub release, the release-notes index at docs.openclaw.ai and the tracker on September 5.

September 8, 2026 - the fifth release in nine days, v2026.9.3, and the first with a breaking runtime requirement. Published September 8 at 14:15 UTC on GitHub, three days after 9.2. By the team’s count: 1,844 pull requests, 40 direct commits, 190 contributors, the largest release since 2.0 itself. The release notes lead with recovery: updates that rehearse core and plugin changes in isolated candidate state before activating them, eligible 9.2 migrations supported, and abandoned update records recovered without stopping a healthy Gateway. That closes the arc the day-one reports opened on August 31.

Read the breaking section before you upgrade, because this one is not a plugin-author footnote:

• Node 24.16.0 or newer on 24.x, or Node 26.1.0 or newer, is now required, with Node 26 recommended. Node 22, Node 25 and earlier 24.x and 26.x builds are no longer supported, and the notes say to upgrade Node before OpenClaw to prevent SQLite text truncation. Node-based installs on macOS 11 to 13.4 and the official Linux ARMv7 path need a supported host. If your Gateway runs on a distro’s packaged Node, this is the release that makes you check.
• Plugin SDK helpers are retired outright: the execution-policy helpers move to execPolicy on the agent-harness runtime, approval account-resolution moves to approval-native-runtime, two aliases are replaced, and Find/Grep and directory-listing result callbacks read different fields. The SDK migration page is now mandatory reading for anyone shipping a plugin.

What is new for operators and agent builders: Skill Workshop keeps each agent’s learned skills in one persistent, agent-owned collection across workspaces, with full-instruction comparison and safe retirement of stale drafts through Doctor. That is the closest OpenClaw has come to the per-agent institutional memory we described as the category’s missing piece. Alongside it: warm prompt caches preserved across sessions and worker builds reused between them; browser automation you can watch repaint live, with external links opening in native Mac tabs that stay with their window; provider accounts and account priority managed in Models settings; shareable sessions (an explicitly published, revocable, read-only public view of a conversation, including future text); a searchable meeting library with full-transcript search and Markdown or JSONL export; optional Team Reports over GitHub activity and configured Discord discussion; and repository-backed cloud sessions created from a URL and ref, with checkpoints owned by the cloud node.

Against the day-one list: the Gemini-embeddings memory-sync abort (134015) is not named in the notes for a fifth release, so the batch-size workaround still stands; the plugin-consent issue (134076) remains closed since September 1. No new Astra or Fable model changes in this release. Verified against the GitHub release and the release notes at docs.openclaw.ai on September 8.

September 11, 2026 - the sixth release in twelve days, v2026.9.4. Published September 11 at 03:46 UTC on GitHub, three days after 9.3. By the release’s own audited record: 1,174 unique pull requests; the notes credit contributors line by line but publish no total. The release notes sum it up as a release that “makes plugins and skills easier to find, and lets you turn past conversations into reusable skills through a chat you can steer.” There is no breaking section this time, and the Node story moves from requirement to repair: the CLI now offers a Node.js update when it finds an incompatible runtime, and keeps its diagnostics and update commands available, instead of refusing to start.

Three things stand out:

• Rollback from a failed update. The Gateway retains the previous package and restores it, with its configuration and service, when schema and configuration checks prove the rollback safe; database migrations still require a verified pre-update backup, and a recovered update is still recorded as a failed one. This is the last step of the arc the day-one upgrade reports opened on August 31.
• Plugins in one workspace. Bundled and ClawHub plugins are discovered, installed and configured from a single Plugins workspace in the Control UI, with one search across installed and ClawHub skills.
• Prepared cloud sessions. Eligible Linux sessions start from prepared local projects or public GitHub repositories, with reusable snapshots built from the Control UI before a conversation begins. Read the reserve settings before turning it on: ready workers “incur provider running-machine charges until deleted”, with a default reserve of one per project and a cap of four Gateway-wide.

Also in: GPT Image 2.5’s Flare and Sunburst variants as image models through OpenAI or fal, keyboard-driven questions in the terminal, and an OPENCLAW_CONFIG_READONLY=1 switch for deployment-managed configuration. The release verification section is unusually candid: the Telegram and Parallels checks were “explicitly waived by the release owner, not passed”, Android native qualification failed, and ClawHub recovery is listed as pending. Against the day-one list: the Gemini-embeddings memory-sync abort (134015) is not named in the notes for a sixth release. No Astra or Fable model changes in this release. Verified against the GitHub release and the release notes at docs.openclaw.ai on September 11.