OpenClaw v2026.10.1: skill search, MCP Apps and the upgrade steps to run first

OpenClaw v2026.10.1 was tagged on the project's GitHub repository on 10 October 2026, and newer releases have overtaken it since: it no longer carries the Latest badge there, and the documentation's release-notes index lists later versions above it, so check the OpenClaw releases page for the version that is current when you read this. Three features lead v2026.10.1: the agent can search the skills it already has installed rather than carry the whole library in its prompt, MCP Apps can run beside chat with access to workspace files, and full backups can be written to an external drive or Cloudflare R2. The part to read before you type an update command is the list of installations that need an intermediate release first, because several older data and configuration formats stop the upgrade instead of converting in place, and a release later than this one may add steps of its own.

One sourcing note, because it affects how you check this for yourself. The release notes page in the OpenClaw docs carries no publication date at all. The dated record is the release on the repository, which shows 10 Oct 23:54 beside the v2026.10.1 tag. If you want to know how old your copy of the docs is, the repository is the page that tells you.

#What is new in OpenClaw v2026.10.1?

The release notes lead with three things, in their own words: the release "helps your agent find a relevant installed skill when a task needs it, without putting the whole skill library in the prompt", then "You can then work with compatible, opt-in MCP Apps beside chat, using supported forms and workspace files", and "To keep your setup recoverable, full backups can now live on external storage or Cloudflare R2 and restore into a fresh staging directory."

Feature	What it changes	What it needs
Installed-skill discovery	The agent searches installed skills for a task instead of loading the whole library into the conversation	Supported embedded agents, Code Mode and Codex sessions
MCP Apps beside chat	An app runs in a pane next to chat, can take attached content, and can view or edit the workspace files it advertises	MCP Apps enabled and a compatible server configured
Backups on external storage	Full backups can live on an external drive, mounted storage or Cloudflare R2, and restore into a fresh staging directory before activation	Scheduled backups are opt-in; a passphrase for encrypted uploads, or an explicit choice of none
Code Mode working values	Small JSON results can be saved and reused in later cells, follow-up replies and after a restart	Interactive Code Mode; only completed cells save their data
Conversation snooze	An eligible conversation leaves the Active sidebar while its work continues, and returns at a chosen time or earlier	The web Control UI; archived, child and protected conversations are excluded
Bundled desktop setup	A fresh packaged Mac app can run OpenClaw through This Mac without installing Node or a separate CLI	The macOS 15 minimum still applies; background running is behind a setting

#What do you have to do before you upgrade?

This is the half of the release notes worth reading twice. Older formats do not silently convert: the notes say "Very old configurations and saved-data formats may need an intermediate release before this upgrade can proceed; refusals preserve the old files." A refusal is the good outcome there, because it leaves your data where it was. The same passage says to "Back up complete configuration, state and queue-owned media first."

If your installation has	Do this first
Workboard KV data from before July 2026	"Run doctor --fix on OpenClaw 2026.9.7 before upgrading". Current Doctor preserves that data and warns rather than importing it
Standalone Teams or Nextcloud Talk webhook plugins you replace by hand	"Upgrade core to at least 2026.9.9 before manually replacing these standalone plugins"
Nextcloud Talk still using the retired private-network setting	The documented "2026.9.5 intermediate upgrade first"
Retired configuration forms: old agent runtime, sandbox, Browser, queue, webchat and authored plugin-install formats	"The documented 2026.9.5 intermediate repair before upgrading"
Retired outgoing or session JSON queues, cron files, encrypted OAuth sidecars, Voice Call logs or nested workspace setup files	"2026.9.7 on the original host, then Doctor"
A manual binary replacement over a legacy agent roster	"Run openclaw doctor --fix before startup". Normal updates already run Doctor through their backup flow

Two cautions from the same notes are easy to miss. Rolling back after the Workboard migration is not clean: "If you roll back, older releases show Sessions boards as empty Cards boards; cards added there are rejected when you return." And Doctor's housekeeping now removes things you may have been treating as archives, on narrower terms than a glance suggests: "After a successful fresh capture, Doctor can remove eligible completed standalone copies older than 30 days, so keep lasting backups separately." The qualifier and the condition are both the page's own, and its per-change entry repeats them while putting the advice as "use a verified backup for lasting recovery". So keep any backup you actually rely on somewhere Doctor does not manage. Our OpenClaw migration guide covers the general shape of a version move, and OpenClaw troubleshooting covers what to do when an update stops.

#What changed for skills and MCP servers?

Installed-skill discovery is the headline, and its limits are documented rather than implied. Search "returns matching names and descriptions, while an exact-name read returns complete instructions or an explicit size error", and instruction reads "support files up to 256 KiB". Searching inside instruction text is partial by design: coverage runs to "up to 16 KiB per file within the total search budget", and words past that prefix are not matched.

Two sentences in that section deserve to travel with the feature. The first is about permissions: "Searching instruction text follows the agent's current read permission and may cover only part of a file, so no match does not mean no suitable skill exists." The second is about what a read does and does not grant: "The agent reads a selected skill's full instructions separately; that read grants no permission to install or run it." So discovery widens what the agent can find, not what it can do, and a no-match answer is not evidence that you have no skill for the job. If you are weighing skills you did not write yourself, our skill security audit is the process to run before any of this matters.

MCP Apps are opt-in twice over, needing both MCP Apps enabled and a compatible server configured. Approval is visible where the action happens: "Messages and file opens show their approval controls in the app pane, with temporary approval available for repeated use of one app tool." The risk the notes name themselves is concurrent writes, since "File editors can overwrite a newer edit if they do not check for stale saves." Treat an app holding write access to your workspace as another writer in the room rather than a viewer.

#What should you do today?

• If you run OpenClaw on anything you care about: take the full backup the notes ask for before updating, and read the prerequisite table above against your own installation rather than assuming the update will sort it out. The refusals are protective, but they still stop you mid-upgrade.
• If you use Workboard: check whether your KV data predates July 2026, and if it does, run doctor --fix on 2026.9.7 before you move. Plan on not rolling back afterwards.
• If your channels run through Teams or Nextcloud Talk webhooks: get core to 2026.9.9 or newer first, move and verify the callbacks before you delete the old listener settings, and keep meta.migrations.webhookListeners.
• If you relied on Fast or Auto reaching Ultrafast: that no longer happens on its own. Choose it with /fast ultrafast or in the picker, and do not treat the choice as a guarantee, because the notes say providers may return a different tier.
• If you want the new skill search: it lands in supported embedded agents, Code Mode and Codex sessions, and it is bounded by the agent's existing read permission, so review that permission rather than expecting discovery to reach everything.

The full feature list is longer than any summary of it, and the per-change entries on the v2026.10.1 release notes carry the pull request numbers if you need to trace a specific behaviour. For what OpenClaw is and what it runs on, start at our OpenClaw hub. Everything above was checked against the release notes and the repository release on 11 October 2026 and re-checked on 22 October 2026, and the release-dating details in the opening and the FAQ were re-checked on 23 October 2026.