OpenClaw v2026.10.1 ships on-demand skill discovery, MCP Apps and full backups
OpenClaw v2026.10.1 (Oct 10): agents search installed skills on demand, MCP Apps run beside chat, full backups to drives or R2 — 10.2 followed in ~30 hours.
OpenClaw is an open-source personal AI assistant — MIT-licensed, around 392k GitHub stars as of October 12, 2026, with a repo tagline that reads “The AI that really does things.” On October 10 at 23:54 UTC it shipped v2026.10.1, a calendar-versioned release whose headline, per the official notes, is installed-skill discovery, MCP Apps and full backups, alongside routine updates to chat, native apps and automations. The scale is checkable: the GitHub release page credits 3,055 commits, 2,425 pull requests and 402 contributors. About 30 hours later, v2026.10.2 took over at 05:34 UTC on October 12.
Key points
- Skills on demand: in supported embedded agents, Code Mode and Codex sessions, the agent searches the installed-skill index according to its read permissions and reads full instructions by name — 256 KiB per instruction read, 8 MiB across a combined set.
- MCP Apps with approvals: apps run beside the chat, selected content attaches to the next message, and approval controls sit in the app pane per tool; a single tool can request temporary approval for repeated use.
- Full backups: destinations include an external drive, mounted storage or Cloudflare R2; a restore unpacks into a fresh staging directory before activation, with an archive-size plus 256 MiB free-space check before extraction, and optional passphrase encryption.
Background
OpenClaw version numbers follow a year.month.patch calendar, and the release density is part of the story. The tag timeline on GitHub (queried October 12) shows nine tags between October 2 and October 12: four stable releases (v2026.8.34, v2026.8.35, v2026.9.8, v2026.9.9), two v2026.10.1 betas, then v2026.10.1 itself on October 10, v2026.10.5-beta.1 the next day, and v2026.10.2 in the early hours of October 12. Pedaily counted a “five releases in five days” streak back in April — different circumstances, similar tempo. Peers are moving fast too: Pi, a terminal coding-agent framework, hit 1.0 on October 1, and the engineering race around personal agents is clearly accelerating.
The facts
The official release notes (docs.openclaw.ai, checked October 12) spell out each feature’s boundaries, numbers included:
| Feature | What it does | Limits | PRs |
|---|---|---|---|
| Skill discovery | In supported embedded agents, Code Mode and Codex sessions, the agent searches the installed-skill index by read permissions and reads instructions by name | 256 KiB per instruction read, 8 MiB combined; oversize reads fail with an explicit error | #158089–#164012 |
| MCP Apps | Apps open beside the chat; selected content attaches to the next message; approval controls live in the app pane | Per-tool approval; one tool can request temporary approval for reuse | #161747 et al. |
| Full backups | Write to an external drive, mounted storage or Cloudflare R2; opt-in schedule with a retention policy, optional passphrase encryption | Restores unpack into a fresh staging directory before activation; extraction is preceded by an archive-size plus 256 MiB free-space check | #161913 et al. |
The skill-search section states the index is filtered by the agent’s read permissions, and an exact-name read either returns the complete instructions or a clear size error. The backup section documents the restore path: a pre-extraction check on archive size plus a 256 MiB free-space reserve, then unpack into a fresh staging directory before activation. Quieter engineering changes ride along in the same release — supported Bun-only installs on macOS and Linux run the plain openclaw command without Node, npm installs save disk by dropping native binaries for unrelated platforms, the optional Workboard plugin organizes conversations on Sessions boards, conversations can be snoozed, and model discovery gets an update.
Approval flows are appearing elsewhere in the MCP world too. When Shopify opened checkout to in-browser AI agents in late September, its three WebMCP tools required per-item buyer authorization; Google Cloud’s enterprise Gemini Agent lists “connects to any MCP server” as a core capability. Building “the user says yes” into the protocol layer is becoming the ecosystem’s default.
What others say
The cost of this pace lands on users first: 2,425 pull requests and 402 contributors packed into one release, stable versions a day or two apart, and by the time a reader finishes the 10.1 notes, 10.2 is already sitting on the releases page. The project is frank about it — the release notes include an “Installed updater limits” section stating that several fixes depend on the new updater being installed first, and that an older updater may refuse mid-operation, at which point the manual update procedure applies with a verified backup.
The mitigating reading comes from the same cadence: v2026.10.5-beta.1 landed the day after 10.1, so mainline changes are still merging at speed. For anyone running OpenClaw in production, the practical approach is closer to “wait a patch” — the roughly 30-hour gap to 10.2 is the ready-made example.
Our take
For self-hosters, the structural change is skill retrieval: instructions move from “always in the prompt” to “searched at runtime by permission, read by name,” which cuts context bloat and token spend, and the 256 KiB / 8 MiB caps signal that the project expects skill collections to keep growing. Per-tool MCP approvals and staging-based restores are both overdue production hygiene for personal agents. The honest catch is timing: if you are upgrading today, go straight to v2026.10.2 and use the 10.1 notes to judge whether the direction is worth following. As for the cost of keeping up — nine tags in eleven days means reading notes and running upgrades on the project’s schedule, and whether that bill is worth paying depends on how much you lean on what OpenClaw does.