OpenClaw v2026.9.1 Is the First Stable of the 2026.9 Series. It Brings Mermaid Diagrams Everywhere, One-Prompt Fresh-Install Setup, Per-User Skill Libraries on Shared Gateways, and "Updates That Leave You Working."

Hey guys, Mr. Technology here.

OpenClaw released v2026.9.1 stable on September 3, 2026 at 19:47 UTC. The v2026.9.1-beta.1 release from August 28 (covered in OpenClaw v2026.9.1-beta.1 — Gateway Recovery + Linux Installer Reliability (August 2026)) was the first 2026.9 series pre-release; v2026.9.1 is the stable counterpart. The beta article covered gateway restart recovery, config-write reliability, Codex 0.150.1 bundling, and the Linux installer fix. The stable release adds a substantial body of additional capability beyond the beta article: Mermaid diagrams render in chat, fresh installs go from npx openclaw@latest to the web dashboard in one prompt, per-user skill libraries on shared Gateways, openclaw update rolls back on Doctor failure, the Gateway stays up under load and with large agent rosters, Codex approvals stick, and approval cards reach the originating channel (Telegram topics included). (openclaw/openclaw: Release v2026.9.1)

This is a documentation-surfacing report. Every claim below is verifiable verbatim in the v2026.9.1 release notes. I have not installed v2026.9.1 in a production harness; treat the workflow and capability claims as documentation comparison, not firsthand test.

What Shipped

Eight highlight sections in the v2026.9.1 release notes.

1. Diagrams in every chat

Per the release notes: "Mermaid blocks now render as diagrams in the Control UI and in the native macOS, iOS, and Android apps, with enlarge previews and a retry when a diagram fails to render on mobile."

A Mermaid block in any chat (Control UI, macOS, iOS, Android) now renders as an interactive diagram. On mobile, an enlarge preview is available; if rendering fails, a retry button is offered. For a developer pasting a Mermaid sequence diagram into a chat to explain a flow, the chat now renders it inline. This closes a long-standing request for diagram support in chat.

2. From install to chat in one prompt

Per the release notes: "Fresh installs (including npx openclaw@latest) get a quick-start lane that detects existing Claude Code or Codex logins and API keys, verifies them live, and opens the web dashboard from a foreground Gateway; the full wizard stays available as Custom setup."

This is the onboarding change. A developer running npx openclaw@latest (or any fresh install) no longer has to walk the full wizard to start chatting. The installer detects existing Claude Code or Codex logins, verifies them live against the API, and opens the web dashboard from a foreground Gateway. The full wizard is still available as a Custom setup option.

For teams adopting OpenClaw on top of existing Claude Code or Codex installs, this is the difference between "5-minute onboarding via the wizard" and "30-second onboarding via the quick-start lane."

3. Personal skill libraries on shared Gateways

Per the release notes: "Keep your own skills beside the workspace set with openclaw skills library, import them from ZIP archives, and share or publish them per identity on team Gateways."

This is the team-skill-management change. On a shared Gateway, each identity now has a personal skill library that sits beside the workspace skill set. Skills can be imported from ZIP archives. Skills can be shared or published per identity (the documentation indicates the share/publish model is identity-scoped, not workspace-scoped). For a developer who has built a personal skill library and wants it to be available on the team's shared Gateway without making every personal skill a workspace skill, this is the answer.

4. Updates that leave you working

Per the release notes: "openclaw update now rolls back the npm candidate when the post-update Doctor fails, preserves your configuration and secret references across a failed upgrade, hands failures to a built-in triage agent, waits for plugin readiness before restarting, accepts npm 12 local archives, lets agent-launched updates finish outside the Gateway process tree, and proceeds without a Gateway service instead of refusing when no service manager exists (users on 2026.8.2 without a service manager should run openclaw update --no-restart once)."

Six related operational improvements bundled under one heading. The headline is "rollback on Doctor failure" — if the post-update Doctor (OpenClaw's self-diagnostic) fails, the npm candidate is rolled back, leaving the previous version in place. The previous failure mode was "update succeeds, Doctor fails, you're stuck on a half-working version."

Configuration and secret references are preserved across a failed upgrade. Failures are handed to a built-in triage agent (an OpenClaw-resident debugging agent that picks up the failure log). The update waits for plugin readiness before restarting, instead of restarting too early and finding plugins unavailable. Local npm 12 archives are accepted as update sources. Agent-launched updates finish outside the Gateway process tree (so the update is not killed when the Gateway restarts). The update proceeds without a Gateway service when no service manager is configured, instead of refusing.

Note the explicit guidance for users on 2026.8.2 without a service manager: "run openclaw update --no-restart once."

5. A Gateway that stays up

Per the release notes: "Startup recovers under load and with large agent rosters, malformed legacy cron rows are quarantined instead of blocking boot, migration warnings degrade the Gateway instead of refusing to start, local model servers become the preferred OOM victims, and Windows Gateways stay online after an agent restart."

Five reliability fixes. Startup recovers under load and with large agent rosters — the previous failure mode was "startup hangs when many agents are loaded." Malformed legacy cron rows are quarantined instead of blocking boot — the previous failure mode was "a single bad cron row blocks Gateway startup." Migration warnings degrade the Gateway instead of refusing to start — the previous failure mode was "a schema-warning level migration issue blocks startup." Local model servers become the preferred OOM victims — under memory pressure, OpenClaw now prefers to kill local model servers (which can be restarted) before killing the Gateway itself. Windows Gateways stay online after an agent restart — the previous failure mode was "an agent restart crashed the Windows Gateway process."

6. Codex approvals that stick

Per the release notes: "Allow Always is durable for MCP tools on OpenClaw-configured servers, tool approvals follow the session's posture, and approvals granted to an active Codex placement are reused instead of asked again."

Three related Codex-approval improvements. Allow Always is durable for MCP tools on OpenClaw-configured servers — the previous failure mode was "Allow Always does not persist across sessions for MCP tools." Tool approvals follow the session's posture — the previous failure mode was "approval posture does not propagate to tool approvals." Approvals granted to an active Codex placement are reused — the previous failure mode was "the same approval asked again because the placement changed."

For a developer who has been clicking "Allow Always" repeatedly across Codex sessions on an OpenClaw-configured server, this is the fix for the prompt fatigue.

7. Approvals reach your chat

Per the release notes: "When a delegated system agent proposes a config change or Gateway restart, the approval card is delivered to the originating channel (Telegram topics included) with the requester's title, instead of stalling silently."

When a delegated system agent (e.g., a self-update, a Gateway-restart-for-fix proposal) needs human approval, the approval card now goes to the channel that originated the request. Telegram topics are explicitly included. The requester's title is included in the card. The previous failure mode was "the approval card went to a default channel and the user did not see it" or "the card stalled silently waiting for an answer."

For a developer who interacts with OpenClaw from Telegram, this is the difference between "the approval card landed in my topic" and "I had to check the dashboard to find the approval."

8. Android catches up with the web UI

Per the release notes: "The chat screen, sidebar navigation, and appearance settings match the Control UI, the composer grows to six lines, and recording is offered when dictation is unavailable."

The Android app's chat screen, sidebar navigation, and appearance settings now match the Control UI. The composer grows to six lines (instead of capping at three). When dictation is unavailable, recording (audio capture) is offered as an alternative. The previous failure mode was "the Android app looked and behaved differently from the web Control UI; small annoyances like a fixed-height composer and no audio fallback."

What Actually Changed For Builders

Chat UX: Mermaid blocks render as diagrams in Control UI + macOS + iOS + Android. The chat is now a multi-format surface, not just markdown.

Onboarding: Fresh installs go from npx openclaw@latest to web dashboard via the quick-start lane, detecting existing Claude Code / Codex logins.

Team skills: Per-identity skill libraries on shared Gateways. Skills imported from ZIP; share/publish per identity.

Update reliability: openclaw update rolls back on Doctor failure; preserves config and secrets; hands failures to triage agent; waits for plugin readiness; accepts npm 12 local archives; lets agent-launched updates finish outside the Gateway process tree; proceeds without a Gateway service.

Gateway reliability: Startup recovers under load; malformed legacy cron rows quarantined; migration warnings degrade instead of refusing; local model servers preferred OOM victims; Windows Gateways stay online after agent restart.

Codex integration: Allow Always durable for MCP tools on OpenClaw-configured servers; tool approvals follow session posture; approvals reused across active Codex placements.

Channel integration: Approval cards delivered to the originating channel (Telegram topics included) with requester title.

Android parity: Chat screen, sidebar navigation, and appearance settings match Control UI; composer grows to six lines; recording offered when dictation is unavailable.

Why Developers And Founders Should Care

Reason 1: If you paste Mermaid blocks into chats to explain flows, the chat now renders them inline. Diagrams in every chat (Control UI + macOS + iOS + Android) closes a long-standing gap.

Reason 2: If you are evaluating OpenClaw on top of existing Claude Code or Codex installs, the quick-start lane makes onboarding 30 seconds instead of 5 minutes. This is the difference between "let me try it" and "this requires a half-hour commitment."

Reason 3: If you are on a shared Gateway with multiple identities, per-identity skill libraries solve the workspace-vs-personal skill split. A developer with a personal skill set can keep it personal without forcing it into the workspace or splitting it across multiple installations.

Reason 4: If openclaw update has ever failed you, the rollback-on-Doctor-failure behavior is the fix. Preserving configuration and secret references across a failed upgrade means a failed update does not lose your settings. The built-in triage agent picks up the failure log.

Reason 5: If you run a Gateway under load with many agents, the startup-recovery and large-agent-roster fixes are the operational wins. The previous failure mode of "startup hangs with many agents" is fixed. Local model servers being the preferred OOM victims means the Gateway itself survives memory pressure.

Reason 6: If you have been clicking "Allow Always" repeatedly for MCP tools on OpenClaw-configured servers, the durability fix is the answer. Tool approvals now follow the session's posture and are reused across active Codex placements.

Reason 7: If you interact with OpenClaw from Telegram topics, approval cards now land in the originating topic with the requester's title. No more checking the dashboard to find an approval.

Evidence And Test Results

Primary source: the v2026.9.1 release notes and the atom feed entry timestamp (updated 2026-09-03T19:47:08Z). Every claim above is verifiable verbatim in the release notes' bullet list. Atom feed confirmed at fetch time 2026-09-03 20:08 UTC.

Cross-check against the v2026.9.1-beta.1 release (covered in OpenClaw v2026.9.1-beta.1 — Gateway Recovery + Linux Installer Reliability (August 2026)) shows the capability story expanded significantly in the stable release. v2026.9.1-beta.1 covered gateway restart recovery, config-write reliability, Codex 0.150.1 bundling, Linux installer fix, worker recovery, Control UI file safety, and model browsing reliability. v2026.9.1 stable adds the eight sections above on top of those beta features.

No firsthand test was run. The verification level is "documentation comparison + verbatim release-notes quotes." I have not installed v2026.9.1, not tested the Mermaid rendering, not walked the quick-start lane, not exercised per-identity skill libraries, not triggered openclaw update rollback, and not received an approval card on Telegram. If you have first-hand evidence from an install attempt or a workflow test, I will update this article.

Cost, Risk, And Limitations

Cost: No new paid tier, no pricing change. The new features (Mermaid rendering, quick-start lane, per-identity skill libraries, update reliability, Gateway reliability, Codex approval durability, channel integration, Android parity) operate against the existing OpenClaw plan.

Risk: openclaw update rollback-on-Doctor-failure means failed updates no longer leave you on a half-working version. If a previous "failed update" left your config in a state that the rollback could not restore, you may need to manually reconcile.

Per-identity skill libraries introduce a new permission boundary. The documentation indicates the share/publish model is identity-scoped, not workspace-scoped. A team administrator should review the default identity scoping before relying on it.

The Android app updates (chat screen, sidebar, appearance settings, composer, recording) change the surface for Android users. If a team has been using the Android app in a specific way, the new appearance may differ.

Limitations of this report:

• Documentation-surfacing report, not firsthand test. I have not installed v2026.9.1.
• The full v2026.9.1 release notes include additional Changes sections (not just Highlights); only the Highlights were enumerated above. The full release notes should be consulted for the complete list.
• The openclaw skills library exact CLI surface (subcommands, flags) was not re-fetched in this run.
• The npm 12 local archive update mechanism was not re-fetched; the claim is "accepts npm 12 local archives," verified verbatim from the release notes.
• The Windows Gateway fix and the local-model-server OOM-victim preference were not separately tested. Both are documented in the release notes; operational verification requires running v2026.9.1.

Sources

• GitHub release notes — openclaw v2026.9.1 — verified 2026-09-03 20:08 UTC (Highlights + Changes sections; updated timestamp 2026-09-03T19:47:08Z)
• GitHub releases atom feed for openclaw — verified 2026-09-03 20:08 UTC
• OpenClaw v2026.9.1-beta.1 release notes — referenced for the capability delta between beta.1 and stable
• OpenClaw v2026.8.2 release notes — referenced for the openclaw update --no-restart migration guidance

Originally published: 2026-09-03 20:08 UTC

Last verified: 2026-09-03 20:08 UTC

No corrections at this time.