OpenClaw v2026.10.1 Released: Smarter Skill Discovery, MCP Apps in Chat, and Backups You Can Actually Restore
OpenClaw shipped v2026.10.1 on October 10, its first stable release since 2026.9.8, and the headline is control: the open-source agent platform can now find the right skill for a task without dragging the whole library into your prompt window, run compatible MCP apps beside a conversation, and keep full backups on storage you own. The release tag is GPG-signed, and the review checkpoint behind it counted 2,391 pull requests from 358 contributors.
Skill discovery without prompt bloat
The marquee feature is installed-skill search. When a task needs a capability, the agent can now search the skills already installed on your machine and surface the relevant one — instead of loading the entire skill catalog into the conversation. The search respects the agent's current read permissions, and selecting a skill grants no authority to install or run it; the agent still reads the full instructions separately before acting. The feature works in supported embedded agents, Code Mode, and Codex sessions.
For anyone running a large personal skill library, this is the difference between an agent that greps through a haystack and one that walks straight to the right shelf.
MCP Apps move into the conversation
v2026.10.1 also expands MCP Apps — interactive tools that run in a panel beside the chat. With a compatible server configured, you can work in an app, select content, and attach it to your next message. Supported apps can now view or edit workspace files, offer richer forms with previews and resource choices, and keep calling their tools while their views stay open, with inline message confirmations.
Approval controls surface in the app pane itself, with a temporary-approval option for repeated use of a single app tool. The project notes the boundaries plainly: file editors can overwrite a newer edit if they don't check for stale saves, and native Codex forms require a prompting approval policy.
Backups that survive bad days
The third pillar is recovery. Full backups can now live on an external drive, mounted storage, or Cloudflare R2, and restore into a fresh staging directory before activation — so a bad update doesn't take the old install down with it. Scheduled backups are opt-in, with optional retention that prunes older copies after a successful backup. Uploads can be passphrase-encrypted or explicitly plaintext, and --output keeps a local archive even when the upload is encrypted.
Under the hood: a reliability sweep
Between the betas (10.1-beta.1 on October 5, beta.2 on October 7) and the stable tag, the release accumulated a broad reliability pass: sessions preserve usage across registry changes and survive queued cancellations; local videos regain inline playback; Telegram progress updates stop flashing preview cards; cloud workers surface their real failure reasons instead of stranding; and Windows worktree creation, Linux Chromium discovery, and ARM64 Playwright startup all got fixes.
The changes list adds Claude Haiku 5.5 support, Code Mode store/load values that survive Gateway restarts, Workboard sidebar navigation with board pinning, inline YouTube playback in the Control UI, sealed immutable installation generations with rollback, and cloud-worker support for enterprise repositories. macOS users get a bundled-runtime setup that runs OpenClaw without installing Node; Linux companions move to bundled Bun.
One editorial caveat from the project itself: the published notes cover a frozen review checkpoint, and final reconciliation "remains pending" — so the PR counts may shift in the audited record. Platform teams should also note the long deprecation ledger in the notes: eleven plugin-SDK behaviors began emitting warnings between September 26 and October 3, with removal targets mostly set for October 1 and beyond.
Why it matters
The through-line of 2026.10.1 is trust infrastructure. Skill discovery cuts the context tax of a growing library; MCP Apps formalize how external tools earn permission inside a chat; external backups make recovery a first-class path rather than a hoping-you-saved ritual. For teams running agents unattended across Telegram, Discord, and Slack — the platforms OpenClaw bridges — those three moves target exactly the failure modes that erode confidence in production agent deployments: wrong capability, unapproved access, and unrecoverable state.
Upgrade guidance: the release is on the project's documented fast path, and the docs recommend reviewing the migration and backup guides before switching. As always with a release this size — 2,391 PRs at checkpoint — staged rollouts beat big-bang upgrades.
Source: OpenClaw official release notes (docs.openclaw.ai), GitHub release v2026.10.1
