AGENTS & WORKFLOWS

OpenClaw 2.0: Read the Migration Alongside the Features

Open Source Factory · Sep 01, 2026 · 6 min read

OpenClaw released v2026.8.1 on August 31 under the 2.0 name. The project reports 933 contributors, including 569 first-timers, and more than 16,000 PRs after a seven-week rebuild. The practical changes concern shared workspaces, credentials and automation approvals.

A rebuilt interface is only part of an OpenClaw upgrade. Existing credentials, shared sessions and scheduled work still have to behave correctly after migration. The versioned release notes are a better guide to that job than the total number of merged pull requests.

In this article

1. The release gap and rebuild
2. Guided onboarding
3. The Control UI changes
4. Shared cloud-session permissions
5. Documented security changes
6. Memory behavior and opt-in
7. Breaking changes in the upgrade
8. Other user-facing changes

The release gap and rebuild

OpenClaw’s release notes describe a rebuild after 49 days without a release, following a much faster earlier cadence. Version 2026.8.1, presented as 2.0, changes onboarding, the Control UI and shared sessions. This article follows the documented user-facing changes rather than claiming a review of every merged PR.

Fig 1 — Scale. One release = half the project's history. Vendor-reported PR count, redrawn.

Fig 2 — Cadence. From every 2 days to 49 days. That gap is the story.

Fig 3 — Who built it. 933 people, 61% first-time contributors.

Takeaway for you: if you tried OpenClaw months ago and bounced off setup, this is a different product. If you've been running it daily, read the updating guide first — sessions now live in SQLite, and downgrading needs a verified backup.

Guided onboarding

This is the headline the team wanted: reuse what you already have, verify it actually works, chat first and configure later.

• Reuse: guided setup now detects ChatGPT / Claude / Codex sign-ins, existing API keys, and local Ollama / LM Studio models — and prefers stronger local models automatically.
• Verify: the exact model choice must answer before it's saved. Local GGUF via managed llama-server (replaces node-llama-cpp) and Gemma 4 as the RAM-gated default.
• Platform: Mac app can move itself to Applications (so updates + launch at login work), Linux/Unix gets openclaw on PATH without hand-editing shell files, ChromeOS/Crostini guide added. Node 22.22.2.

Fig 4 — Onboarding 2.0 in four steps. The old flow was "configure everything, then chat." This one is flipped.

Fig 5 — The rebuilt browser Control UI — now a first-class workspace, not a second screen. You can keep setting things up, return to work, or follow live. Image: openclaw.ai.

The Control UI changes

Most people meet OpenClaw in the browser, and the team rebuilt that surface as the primary place to live: durable session progress cards that survive reloads, subagent activity streaming, pinned widgets and dashboards, and structured question cards (buttons + free-text + Skip) that work across web, native apps, and messaging. It's the same session, just visible from wherever you are.

For someone like me who lives in Hermes + browser previews, this is familiar — OpenClaw just decided the web app is the app.

Shared cloud-session permissions

This is the most interesting shift. Before 2.0, you couldn't bring another person into live work without losing what the Claw knew. Now you can run work on paired devices or cloud workers, move the workspace with the session, and hand it off with context intact. Warm machines + project seeds make the next cloud session faster. The team says they used it to build 2.0 itself.

Fig 6 — Shared cloud sessions — "The Daily Claw" dashboard with live teammates. This is the multiplayer workspace they dogfooded. Image: openclaw.ai.

Plain language: your Claw starts simple — watch one inbox, send one Telegram alert — and grows to cross-app tasks ("find that iPad receipt my brother asked about on iMessage") and then to team handoffs. Same relationship, wider reach.

Documented security changes

Security is where 2.0 is most deliberate — and also where the release notes are most careful to say what these controls are not. Quoting the docs: team operator roles are "collaboration features and should not be treated as hostile multi-tenant isolation."

Fig 7 — Five surfaces hardened at once. Not one hotfix — a sweep. Vendor-reported scope, redrawn.

• Credentials: private credential requests via masked prompts (value never hits chat or model context), opt-in proxy that only substitutes secrets to approved destinations, shared team credential store (SQLite, write-only), and an optional 1Password broker with per-secret approval + audit.
• Plugins: install/enable now shows capability, source, version, and exact artifact before you consent. Arbitrary executables need --force. ClawHub audit info surfaced in the flow.
• Sessions: explicit permission modes + workspace restriction anchored to the recorded workspace/worktree. Less chance of wandering outside the approved directory.
• Automations: approve a recurring operation once for the exact operation, inspect/revoke later, and re-approve when the job changes — so an approved workflow can't silently expand.
• Models + config: model allowlists (with per-agent policy + wildcards), config-change history with sensitive-value redaction, DB recovery protections, sanitized triage handoffs.

The credential proxy and per-operation automation approvals limit particular forms of credential exposure and unattended action. They do not establish multi-tenant isolation. A shared workspace needs its documented trust boundary read before it is used as a tenant boundary.

Memory behavior and opt-in

Three related defaults changed:

Feature	Default	What it does
Personal conversation recall (Active Memory)	On for personal installs	Same-agent private context recalled in a bounded way. Groups/channels excluded. You can disable.
Grounded dreaming	On	Background consolidation promotes provenance-qualified material into long-term memory, with a Dream Diary.
Automatic self-learning	On (scanner-approved)	Strong reusable lessons + Workshop-owned skills can auto-apply; your own skill edits stay pending.

All three have an explicit off switch. I like that — memory as a default you can keep, not a surprise you discover.

Breaking changes in the upgrade

Before you update: create a verified backup. Sessions moved to SQLite — sessions created after migration won't appear if you downgrade to a file-backed release. Restore legacy transcript artifacts before downgrading. Also: the package tagged 2026.9.1-beta.1 is misversioned — it's really 2026.8.1-beta.4, not newer than stable.

• OpenProse removed: bundled plugin + /prose command gone. Run openclaw doctor --fix and follow the Agent Skill migration. Your .prose files stay.
• OpenAI routes renamed: codex/* and openai-codex/* → openai/*. openclaw doctor --fix migrates model refs, provider config, stored sessions, and automation routes.
• Provider plugins unbundled: BytePlus, ComfyUI, Mistral, NovitaAI, OpenCode, Synthetic, Volcengine, Vydra, Xiaomi, Cohere, Meta, DuckDuckGo, Voyage, iMessage now install as separate official plugins when needed. Missing owners recovered via openclaw update repair or doctor --fix.
• SDK deprecations Sep 1, 2026: external plugins should migrate to api.pluginConfig / openclaw/plugin-sdk/* focused imports (channel-outbound/inbound, config-contracts, delivery-queue-runtime, etc.). Not removed yet — just the gate.

Other user-facing changes

Find past conversations by exact phrase, session search reopening surrounding messages; audio/video stays attached across uploads, replies, playback and reloads; Discord voice improvements, native video to Google/Moonshot models, Fish Audio + local MLX speech, Feishu meeting join, Linux PipeWire talk-back. Full list is in the release notes — this post covered the shape, not the tail.

The guided setup and Control UI target the earlier onboarding friction. Existing installations also need the breaking-change review, a backup and checks of credentials and automations. Shared sessions add a separate access-control question: who can see the context and initiate an action.

Follow the versioned migration notes, verify the Gateway starts and inspect the existing scheduled work after an update. A clean startup is the beginning of that check, not proof that every old integration still behaves as expected.

Sources: openclaw.ai/blog/openclaw-2-accidentally · docs.openclaw.ai/releases/2026.8.1 · github.com/openclaw/openclaw v2026.8.1 · cybersecuritynews.com — security summary
Images: Fig 5,6 via openclaw.ai (credit in caption); Figs 1-4,7 self-rendered charts. Stars/forks as of Sep 1, 2026.
Related on this blog: Qwen3.8-27B: 27B Dense, Vision Included, Apache-2.0 · 2026 Local AI Machines Ranked by Value: Which Box Actually Earns Its Place

Edited October 26, 2026. Original publication and evidence dates are retained; this edit is not a new product test.

Reading guideAbout & methodsCorrectionsPrivacy