OpenClaw 2026.9.1 adds useful interface features, tighter operational safeguards, and a long list of reliability improvements for self-hosted OpenClaw installations. Released on September 3, 2026, it includes in-chat Mermaid diagrams, personal skill libraries, safer update checks, stronger Gateway behavior, native-app improvements, and more predictable browser automation.
This guide explains the changes that matter most to OpenClaw users, VPS operators, developers, and teams—and gives you a practical checklist for upgrading safely.
OpenClaw911.com is an independent OpenClaw services and support website operated by All About Web Services. It is not the official OpenClaw project.
OpenClaw 2026.9.1 at a glance
- Mermaid diagrams rendered directly in web and native conversations
- A shorter Quick Start installation and onboarding flow
- Private skill libraries for users of a shared, trusted Gateway
- Stronger readiness checks and recovery behavior during updates
- Gateway improvements for large agent rosters and session histories
- A more complete Android chat and session experience
- Clearer model selection and bounded provider fallback behavior
- More predictable browser and computer-use automation
- Safer memory-index reset and more efficient search maintenance
- Security improvements around approvals, diagnostics, and destinations
The official release record credits 281 contributors and covers 1,186 pull requests plus 28 direct commits. That breadth is a good reason to review the sections relevant to your own deployment before updating.
Mermaid diagrams inside conversations
Completed Mermaid code blocks can now render as visual diagrams directly in OpenClaw conversations. The feature is available in the Control UI and the native Android, iPhone, iPad, and Mac applications.
In the web interface, users can inspect and copy the source, expand the diagram, and zoom in. Invalid or incomplete Mermaid syntax falls back to readable source rather than presenting a misleading visual. The renderer is isolated and lazy-loaded, and its output is converted to passive SVG before display.
This is particularly useful for architecture maps, deployment pipelines, process flows, database relationships, sequence diagrams, troubleshooting paths, and project dependencies.
Faster Quick Start installation and onboarding
Fresh installations now have a shorter route from setup to the web dashboard. Quick Start can detect supported existing Claude Code or Codex access, as well as an existing provider key, verify the selection, and open the dashboard after a simplified setup decision. The full custom wizard remains available when more control is needed.
Headless installations can provide a one-time authenticated dashboard link and an SSH-tunnel template. This reduces the need to expose reusable Gateway credentials just to finish onboarding.
The guided flow also handles cancelled or failed provider checks more carefully, preserving the prior working configuration instead of leaving a partial replacement behind.
Private skill libraries for shared teams
OpenClaw 2026.9.1 introduces personal skill libraries for signed-in users working through a shared Gateway. Team members can create, import, update, and reuse skills without direct access to the host machine.
Libraries start private. Sharing a skill does not automatically grant edit rights, and a skill can later be transferred to team ownership when it becomes shared organizational work. OpenClaw also preserves the selected skill revision—including supporting files and executable permissions—as a session moves between supported execution environments.
Important limitation: this model operates within one trusted Gateway domain. It is not automatic synchronization between independent OpenClaw servers.
Safer updates, readiness checks, and recovery
The update path receives some of the most consequential operational changes in this release. Before restart, OpenClaw performs additional readiness checks against the installation that will actually load. Plugin compatibility issues, unresolved capability consent, and incomplete approval repairs can stop an update before it causes a broken restart.
For supported global npm installations, the updater can restore the previous package and command launchers when the post-update Doctor check rejects the new installation. Managed Linux and macOS Gateways can hand an agent-initiated update to a detached helper, allowing the process to continue after the old Gateway stops.
Update and Doctor workflows are designed to preserve configuration, environment references, workspaces, agent rosters, plugin packages and policies, credentials, scheduled work, device pairings, and session history. When OpenClaw cannot prove that a repair is safe, stopping for operator review is the intended result.
Gateway reliability for larger installations
Gateways with large agent rosters can reuse prepared information during startup instead of repeatedly rediscovering the same agents, models, and workspaces. That helps genuine HTTP health checks answer while remaining startup work continues.
Large session histories also retain less conversation data during lightweight listing and status operations. Read-only diagnostics avoid unnecessarily entering SQLite’s writer lifecycle, while Discord and Telegram activity no longer has to wait for all session-store maintenance before dispatch.
The performance claims are workload-specific. Official testing reports improved headroom in tested 768 MB and 1 GB ARM64 environments, but a 512 MB machine without swap can still fail when status work overlaps startup. Treat these as useful test results—not a universal percentage improvement or compatibility guarantee.
Android, iPhone, iPad, and Mac improvements
The Android app moves closer to the browser workspace with a fuller chat and session interface, improved sidebar navigation, appearance controls, clearer model names, and searchable model, reasoning, Fast Mode, and permission settings.
The composer can grow to six visible lines, narrow layouts and large text are handled more carefully, and reconnect plus queued-message behavior is more reliable. Mermaid diagrams include full-screen and retry controls.
The iPhone, iPad, and Mac apps also gain Mermaid rendering along with improvements to reconnect behavior, conversation state, voice ownership, and background-session completion.
Clearer model selection and provider recovery
Model selection now makes the scope of a change clearer before it is saved: the current session, one agent, or the global default. The Control UI consolidates the primary model, utility model, first fallback, reasoning level, and Fast Mode in one defaults card.
Provider authentication and configuration changes can refresh the model catalog without always requiring a Gateway restart. Temporary provider failures follow a more bounded retry path: a short interruption can recover, while a persistent failure can move to a configured profile or model fallback without marking every candidate as broken.
The release also expands or updates routes for Anthropic Fable 5.1 metadata, SuperGrok usage reporting, local Ollama, and llama.cpp discovery.
More predictable browser and computer use
Standalone agents with browser access now default to the local browser when no Gateway or node route has been selected. This is an intentional behavior change: installations that relied on implicit browser-node discovery should select auto or choose a node explicitly.
Other browser changes include more recoverable screenshot failures, better preservation of mobile viewport and touch settings, validation of form-fill requests before page changes, improved cancellation boundaries for clicks and typing, greater stability for selected Chrome tabs, and hostname-policy checks for custom Chrome MCP endpoints.
Memory reset, indexing, and search
The memory system now separates rebuildable index data more clearly from durable conversations and source files. The memory-reset workflow can remove derived index data while preserving sessions, transcripts, memory source files, and other durable agent state.
Search can reuse a healthy index instead of beginning unnecessary maintenance after it has found an answer. Incremental changes can also be applied without repeatedly rebuilding the complete index.
A memory reset is not a privacy-erasure operation, does not shrink existing SQLite allocation, and cannot recover history that was already deleted. Rebuilding may contact a paid embedding provider, depending on the configured setup.
Automation and scheduling changes
The optional cron.skipMissedJobs: true setting lets recurring jobs that became overdue during downtime advance to their next future run instead of executing every missed occurrence after startup. It is opt-in and does not skip one-time jobs.
Scheduled work also retains the selected agent, model alias, and delivery route more reliably. Invalid legacy schedule records can be quarantined instead of preventing valid jobs—or the Gateway itself—from starting.
Security and privacy improvements
Durable MCP approvals can remain tied to an exact agent, configured server, and tool instead of becoming blanket permission. Stricter session or server policies still take precedence.
Remote Codex placement approvals now have a shorter lifetime. A Gateway restart, placement move, re-pairing, authority change, or expiration can require a fresh approval.
Additional changes include redaction of secret-shaped values in specified diagnostic paths, less sensitive information in deep links, owner-restricted managed Linux backups, stronger authorization during delegated-agent creation, configurable blocking of selected browser, web-fetch, and webhook destinations, patched URI and mail dependencies, bounded A2A request and response sizes, and limits on simultaneous unauthenticated SMS webhook reads.
The hostname blocklist is a destination policy—not a complete network-egress firewall. Continue to use appropriate operating-system, container, and network controls.
Compatibility considerations
- Unpinned standalone browser runs now default to the local browser.
- Chromium or Chrome is required for meme PNG rendering.
- Some private workspace containers may be recreated once for the new mount format.
- Local npm 12 archive paths cannot contain commas.
- Some Plugin SDK compatibility paths are temporary and scheduled for later removal.
- Specific QQ Bot, Weixin, and Voice Call configurations may need migration work.
- An update may intentionally stop when consent, ownership, compatibility, or migration safety cannot be verified.
There is no single universal breaking-changes list for every deployment. Review the official sections that correspond to the integrations and execution environments you actually use.
A practical safe-upgrade checklist
- Review the official release notes and updating guide.
- Confirm your installed plugins and execution environments are supported.
- Check update status and run a dry run.
- Create a verified backup in an owner-only location.
- Run the update and allow readiness checks to stop it if safety cannot be proven.
- Run Doctor, restart the Gateway if required, and check health.
- Verify active plugins, browser integration, message delivery, scheduled work, and at least one real agent response.
openclaw update status
openclaw update --dry-run
openclaw backup create --output <OWNER_ONLY_BACKUP_DIRECTORY> --verify
openclaw update
openclaw doctor
openclaw gateway restart
openclaw healthIf you use the unpacked Chrome extension, also run:
openclaw browser doctor --browser-profile chromeAutomatic configuration copies are not a replacement for a complete recovery point. Backups may contain credentials, authentication profiles, and channel state, so protect them accordingly.
Watch the OpenClaw 2026.9.1 overview
Prefer a concise walkthrough? Watch our video: OpenClaw 2026.9.1: Biggest New Features and Safe Upgrade Guide.
Frequently asked questions
Is OpenClaw 2026.9.1 a stable release?
Yes. Version 2026.9.1 was published as an official stable release on September 3, 2026.
Does it share skills between different servers automatically?
No. The documented personal skill-library feature works within one trusted shared Gateway. Synchronizing skills across independent Gateways requires a separate controlled process.
Will the update preserve existing configuration?
The updater and Doctor preserve more existing configuration and durable state, but operators should still create and verify a complete backup before upgrading.
Does this release improve performance?
Official tests found lower CPU and memory use in several long-conversation, concurrent-streaming, and large-installation workloads. Results depend on the deployment, so avoid applying those findings as a universal performance promise.
Are there security improvements?
Yes. The release adds narrower approvals, diagnostic redaction, backup protections, destination controls, and resource limits. These measures complement rather than replace normal host, container, and network security.
Can an update still fail?
Yes. An update can intentionally stop when OpenClaw cannot verify plugin readiness, capability consent, service ownership, or migration safety. This fail-closed behavior protects the working installation.
Should you upgrade?
OpenClaw 2026.9.1 is especially relevant if you maintain a long-running Gateway, operate large agent or session collections, rely on browser automation, manage reusable team skills, run scheduled jobs, or use OpenClaw on Android or Apple devices.
The release combines visible improvements with deeper maintenance and reliability work. Back up first, review the compatibility notes that apply to your system, and test the workflows your deployment depends on.
Need help with your OpenClaw upgrade?
Contact OpenClaw911.com for independent OpenClaw installation, troubleshooting, compatibility review, backup planning, and post-update health checks from All About Web Services.
