OpenClaw 2026.10.1 lets agents search installed skills and changes how sessions message each other

OpenClaw 2026.10.1, published on October 10, changes what an agent loads before it starts work. Instead of listing every installed skill in the prompt, the agent can search them and read the full instructions of the one it picks. The release also adds opt-in MCP Apps beside chat and backups that live off the machine, and it changes three behaviors that existing workflows may rely on. The GitHub release counts 3,055 commits, 2,425 pull requests, and 402 contributors.

Skills found on demand

A skill in OpenClaw is a packaged set of instructions an agent follows for a kind of task. Until now, eligible skills were listed in the prompt, and the prompt's size limit could leave some of them out. In 2026.10.1, embedded agents, Code Mode, and Codex sessions get tools that search installed skills by name, description, and instruction text. The agent then reads one skill's complete instructions by its exact name.

The release notes spell out the limits. Search follows the agent's current read permission and covers only the first 16 KiB of a long instruction file. A search with no match therefore does not prove that no suitable skill exists. Reading a skill grants no permission to install or run it. Instruction reads now accept files up to 256 KiB.

Apps beside chat and backups off the machine

MCP Apps let a server built on the Model Context Protocol (MCP) show its own interface next to the conversation. With the feature turned on, you can attach content selected in the app to your next message, and supported apps can view or edit workspace files. Approval controls appear inside the app pane, and one app tool can get a temporary approval for repeated use. The notes warn that a file editor that does not check for stale saves can overwrite a newer edit.

Full backups can now be written to an external drive, mounted storage, or Cloudflare R2, and restored into a fresh staging directory before they are activated. Scheduled backups and retention are opt-in. Uploads are encrypted with a passphrase unless you explicitly choose plaintext, and a restore checks for room for the archive plus a 256 MiB free-space reserve before it extracts anything.

What existing setups need to change

The release notes flag three changes that can stop a setup that worked on 2026.9.7.

• Workflows that use sessions_send now get one peer reply per send. Automatic exchanges and posts to a target channel are removed, so a workflow must send again to continue, and it reaches people through message with a channel and a target.
• The earlier required argument is ignored. An agent that should wait for command and child-agent results must use awaitResults:true instead.
• Workboard data stored before July 2026 needs an intermediate step: run doctor --fix on 2026.9.7 before upgrading. The current Doctor keeps that data and warns instead of importing it.

Installation and credentials

Bun-only installations on macOS and Linux can run the openclaw command without Node.js, using Bun 1.4 or newer. npm packages take less disk space because they no longer include native binaries for unrelated platforms. Installations that are already oversized keep their extra files.

Administrators can now let an agent use its managed GitHub identity inside a Docker or Podman sandbox. The setting is per agent and off by default. The notes warn that turning it on exposes the full credentials of the selected profile to sandbox code and background processes. Separately, OpenClaw started through its managed Bun runtime or the bundled macOS runtime no longer installs missing packages on its own, so dependencies have to be installed explicitly.