OpenClaw v2026.9.1: Mermaid Diagrams, Safer Updates, and Better Mobile Chat

OpenClaw v2026.9.1 is a large usability and reliability release. The headline changes are inline Mermaid diagrams, a more complete Android workspace, safer update and rollback behaviour, clearer model selection, stronger session continuity, and lower CPU and memory use in tested long-conversation workloads.

The official release scale is substantial: 1,186 pull requests, 28 direct commits, and 281 contributors. That does not mean every installation should be upgraded blindly. It means operators should identify which changes affect their environment and verify the complete workflow after updating.

Last verified: September 4, 2026, using the official OpenClaw v2026.9.1 release notes. OpenClaw changes quickly, so check the current documentation before modifying a production Gateway.

OpenClaw v2026.9.1 at a glance

The most important changes are:

• Mermaid diagrams render inside web and native conversations.
• Android gains a fuller chat, session, navigation, model, permission, and appearance experience.
• Updates perform more readiness checks before restarting a Gateway.
• Failed global updates can restore the previous package and launchers.
• Model selection explains whether a choice affects one session, one agent, or the global default.
• Provider and model-catalog changes appear without requiring the usual Gateway restart.
• Telegram, Slack, WhatsApp, Matrix, Signal, iMessage, and external-channel paths receive reliability fixes.
• Browser routing and screenshot behaviour become more explicit and recoverable.
• Long histories and concurrent streamed replies use less repeated CPU and memory work in tested workloads.

This is not one isolated feature release. It is an operational-quality update across installation, conversations, native apps, agents, channels, browser control, updates, security, and performance.

Mermaid diagrams now render inside conversations

Completed Mermaid code fences can now render as diagrams directly in the Control UI. The rendered result keeps its source, copy control, expanded view, and zoom controls nearby, so users can inspect both the visual and the underlying definition.

The same capability extends to Android, iPhone, iPad, and Mac. Rendering stays local instead of sending the diagram to an external rendering service. When the source is incomplete, invalid, too large, or temporarily fails, OpenClaw preserves readable Mermaid source rather than replacing it with an empty result.

This is useful for agent-generated:

• system architecture diagrams;
• process and approval flows;
• sequence diagrams;
• deployment topologies;
• data pipelines;
• incident timelines;
• project and task relationships.

The security detail matters. The web renderer reduces model-authored output to passive SVG before display, while native apps provide their own local preview behaviour. Teams should still treat model-generated diagrams as drafts and verify architecture or security claims before using them as documentation.

Android becomes a more complete OpenClaw workspace

The Android app now covers more of the normal chat and session workflow without forcing users back to the browser interface.

The updated experience includes:

• a fuller-width composer with compact actions;
• readable model names;
• searchable model, reasoning-effort, Fast Mode, and permission controls;
• native session browsing and creation;
• configurable sidebar organization;
• theme and accent persistence across reconnects;
• improved narrow-screen and large-text behaviour;
• clearer guidance when an explicitly selected model lacks authentication.

The release also fixes several reconnect and queue-state problems. Accepted messages can move beyond a stale “Sending” state after reconnect, new conversations remain selected, later queued sends continue, and a deleted queued message is less likely to reappear after an older refresh finishes.

For operators, this makes Android more credible as an everyday control surface. It still depends on the connected Gateway’s actual model availability, permissions, and health.

Updates stop earlier when the installation is not ready

The update system now distinguishes more clearly between work OpenClaw can complete safely and conditions that need operator attention.

Before restarting, update readiness checks can catch missing local-memory setup, unresolved plugin consent, incomplete approval migration, and other conditions that would make the next start unsafe. Some nonfatal migration warnings can leave the Gateway reachable in a clearly degraded state so the operator can run Doctor instead of entering a restart loop.

For managed Linux and macOS Gateways, an agent-started update can hand work to a detached helper before the service stops. Windows uses a Gateway-owned restart path so the updater does not disappear with the old process tree.

If Doctor rejects a global npm update, OpenClaw can restore the previous package and launchers while leaving the managed Gateway stopped for inspection. That is safer than claiming success or repeatedly restarting a partially migrated service.

The important limitation is explicit: an accepted update handoff proves that the update was accepted—not that it completed successfully. Post-update verification is still required.

Existing configuration and state receive more protection

OpenClaw v2026.9.1 makes more effort to preserve the setup already authored on the machine, including:

• configuration includes and environment expressions;
• agent rosters and workspaces;
• plugin payloads and policy;
• credentials;
• scheduled work;
• device pairings;
• session history.

When old and new state disagree, repair paths aim to prove that a change is safe or stop before writing. Concurrent Gateway work is also less likely to be overwritten by an older Doctor snapshot.

These safeguards reduce risk, but they do not replace a backup. Persistent state, configuration, credentials references, workspace files, service definitions, reverse-proxy configuration, and rollback instructions should be captured before a major update.

Model and provider selection is clearer

Model selection now tells the user whether a choice applies to the current session, a single agent, or the global default before saving it. That reduces a common operational mistake: changing a model in one interface without understanding which other sessions or agents will inherit the change.

Provider logins and catalog changes can flow into the model list without requiring a Gateway restart. The release also includes work across Codex, Claude Code, local providers, and model authentication paths.

After upgrading, verify the resolved model state rather than trusting the picker alone:

openclaw models status --check

Confirm the primary model, fallback chain, provider authentication, and any allowlist that constrains what the interface can display.

Channel and delivery reliability improves

The release includes a broad collection of channel-specific fixes rather than one universal messaging rewrite.

Notable areas include:

• Telegram recovery from stale session starts;
• faster acknowledgement of durably stored Telegram button presses;
• Slack Socket Mode and relay startup without resolving an unused HTTP secret;
• clearer Slack inbound-rejection diagnostics;
• Matrix sync-database repair through Doctor while preserving state;
• continued inbound reply dispatch for older supported Feishu, Telegram, WhatsApp, and Matrix plugins;
• external-channel compatibility with generic message actions and broadcast planning;
• clearer recovery when Signal or iMessage delivery fails.

Configured channels that fail now remain visible as blocked with a recovery hint instead of disappearing from status and health reporting. That is a meaningful operational improvement because invisible failures are much harder to diagnose.

After the update, probe every channel you depend on:

openclaw channels status --probe

A healthy Gateway does not prove that Telegram, WhatsApp, Slack, or another channel can receive and deliver a real message.

Browser and computer-use behaviour is more predictable

Standalone agents with the browser tool enabled now default to the local browser when no Gateway or node route is selected. Anyone who relied on implicit browser-node discovery should select auto or an explicit node instead.

Screenshot capture now follows the page session that owns the current device settings, including mobile scale and touch behaviour. Interrupted Chromium captures return a recoverable error rather than leaving later screenshots or viewport changes stuck.

Selected Chrome tabs retain valid commands through ordinary group changes, while real access changes can still revoke stale work. OpenClaw also pins existing-session attachment to a reviewed Chrome DevTools MCP release instead of following a moving latest tag.

Because browser routing has an intentional default change, include one safe browser or screenshot task in the post-upgrade test plan.

Lower overhead during long conversations

The streaming path does less repeated work as replies grow. OpenClaw reports substantially lower CPU and memory use in its defined local test workloads for concurrent long replies, while preserving tool order, corrections, reply targets, and final media.

Related changes include lighter session reads, scanning long histories without copying them, reducing repeated subscription work, coalescing obsolete snapshots for slow consumers, and keeping large saved prompts out of lightweight Gateway metadata reads.

These are implementation-level efficiency improvements, not a promise that every deployment will feel equally faster. Model-provider latency, network conditions, hardware, conversation shape, and plugin behaviour still affect real performance.

Does v2026.9.1 contain breaking changes?

The official release page does not provide a dedicated breaking-changes section, but it documents intentional behaviour changes and stricter readiness checks.

Pay particular attention if your setup depends on:

• implicit browser-node discovery;
• external or older channel plugins;
• custom provider onboarding;
• managed service restarts;
• pending plugin consent;
• approval or configuration migrations;
• npm 12 local archives with unusual paths;
• large multi-agent or multi-workspace installations.

Treat “no breaking-changes heading” as different from “no compatibility risk.” Review the official notes for the components you actually operate.

A safe OpenClaw v2026.9.1 upgrade checklist

1. Record the current state. Capture the installed version, active services, Gateway health, models, channels, plugins, agents, and scheduled work.
2. Create and verify a backup. Include persistent databases, configuration, credentials references, workspaces, skills, service definitions, and proxy configuration.
3. Review readiness. Resolve plugin consent, approval migration, local-memory setup, and known Doctor warnings before restarting.
4. Use the supported update path. Follow the current installation-specific instructions instead of mixing npm, source, container, and desktop update methods.
5. Confirm the installed version. An accepted update request is not evidence that the new version is running.
6. Require a healthy Gateway probe. Use openclaw gateway status --require-rpc and inspect logs if it fails.
7. Verify models and channels. Run model authentication checks and channel probes.
8. Reopen existing work. Confirm sessions, histories, agents, workspaces, tools, and scheduled automations remain present.
9. Test changed surfaces. Render one Mermaid diagram, use the relevant native app, and run a safe browser task if those features matter to you.
10. Inspect logs before closing the change. Look for migration, permission, plugin, channel, restart, proxy, or state-path warnings.

Should you upgrade?

OpenClaw v2026.9.1 is especially relevant for teams using the Control UI heavily, working from Android or Apple devices, maintaining multiple sessions or agents, operating messaging channels, or running long concurrent conversations.

The safer update workflow is also valuable for production operators, but it should not be mistaken for automatic acceptance. Back up first, upgrade through one supported path, and verify the real browser, model, channel, and session workflows your users depend on.

If you need help planning or recovering an update, I provide OpenClaw setup and support for installation, model providers, messaging channels, private remote access, troubleshooting, upgrades, and production verification. You can also review my Hermes and OpenClaw deployment work or book a free strategy call.

Official source

• OpenClaw v2026.9.1 release notes