> ## Documentation Index > Fetch the complete documentation index at: https://docs.loongclaw.ai/llms.txt > Use this file to discover all available pages before exploring further. # Skills > Install, expose, and troubleshoot skills without guessing. # Skills Use this page when the question is no longer "does Loong support skills?" and is now "how do I discover, install, expose, and troubleshoot them on purpose?" `loong skills` is the public operator surface for discovery, install, removal, and persisted policy updates around skills. ## Start Here | If you need to... | Start here | | ------------------------------------------- | -------------------------------------------------------------------------------------------- | | inspect the current persisted skills policy | `loong skills policy get` | | see what is already resolved in the runtime | `loong skills list` | | search or recommend by operator goal | `loong skills search "browser automation"` or `loong skills recommend "summarize docs site"` | | inspect one installed or discoverable skill | `loong skills info ` | | install a local directory or archive | `loong skills install ./path/to/skill` | | fetch and install a remote package | `loong skills fetch https://example.com/skill.zip --approve-download --install` | | install a bundled browser automation helper | `loong skills install-bundled agent-browser` | ## Default Public Posture Skills start from a fail-closed default: ```toml theme={null} [skills] enabled = false require_download_approval = true allowed_domains = [] blocked_domains = [] auto_expose_installed = false ``` Rules that matter: * `enabled = false` keeps the skills runtime fail-closed rather than silently widening capability. * `require_download_approval = true` means remote skill downloads require an explicit operator approval step. * `auto_expose_installed = false` means installation and runtime visibility are kept separate until you turn that gate on. * `install_root` is optional and is the right place to pin a stable skill location when the default runtime-owned path is not enough. If you want the broader config map before this page gets operational, start with [Configuration Patterns](/use-loong/configuration-patterns). ## Operator Mental Model Skills have four separate states: 1. discovery: `loong skills list`, `search`, `recommend`, and `info` tell you what the runtime can resolve; they do not install anything 2. installation: `loong skills install` or `fetch --install` copies a skill into the install root 3. exposure: `skills.enabled` and `auto_expose_installed` decide whether installed skills become part of the usable runtime surface 4. runtime readiness: some skills, especially browser automation helpers such as `agent-browser`, still depend on shell policy or an external runtime binary on `PATH` That separation is why "the skill installed correctly" and "the assistant can use it right now" are not always the same statement. ## Core Operator Commands Discovery and inspection: ```bash theme={null} loong skills list loong skills search "browser automation" loong skills recommend "summarize docs site" loong skills info agent-browser ``` Install and remove: ```bash theme={null} loong skills install ./downloads/my-skill loong skills install ./downloads/my-skill.zip --skill-id my-skill --replace loong skills fetch https://example.com/my-skill.zip --approve-download --install loong skills remove my-skill ``` Persisted policy: ```bash theme={null} loong skills policy get loong skills policy set --enabled true --auto-expose-installed true --approve-policy-update loong skills policy set --allow-domain skills.sh --block-domain internal.example --approve-policy-update loong skills policy reset --approve-policy-update ``` Operational notes: * use `skills search` or `skills recommend` when you do not already know the skill id. * use `skills info` after install to inspect prerequisites and follow-up steps. * `skills fetch` is the remote-download path; if approval is required, include `--approve-download`. * `skills policy` persists runtime policy into config; it does not just change one in-memory session. * `skills policy` can toggle enablement, download approval, domain allow/block rules, and `auto_expose_installed`, but `install_root` still belongs in config. * `skills install-bundled ` also exists for first-party packaged skills, including `agent-browser`. ## Local Install vs Remote Fetch Use local install when the package is already on disk: ```bash theme={null} loong skills install ./skills/browser-helper ``` Use remote fetch when the package lives behind a URL and you want the runtime to apply the same domain and approval policy that operators see in config: ```bash theme={null} loong skills fetch https://example.com/browser-helper.zip \ --approve-download \ --install ``` Choose deliberately: * `install` is best for local archives, unpacked directories, or reviewed build artifacts you already trust. * `fetch --install` is best when you want one operator command that both downloads and installs under the governed skills policy. * domain allow/block rules apply to `fetch` and belong in persisted policy rather than being hidden in ad-hoc shell aliases. ## Agent Browser Setup The shortest public route for richer browser automation is to install the bundled `agent-browser` skill and make sure the runtime binary is available. Typical operator loop: ```bash theme={null} loong skills install-bundled agent-browser --replace npm install -g agent-browser && agent-browser install agent-browser open example.com loong doctor ``` Two boundaries still matter: * if `[tools].shell_deny` contains `agent-browser`, the runtime path will stay blocked until you remove that hard deny. * if the runtime binary is still missing on `PATH`, the skill may be installed but automation will not be ready yet. ## Persisted Policy Shape When you want the config to be explicit instead of relying only on CLI mutations, this is the public shape to keep in mind: ```toml theme={null} [skills] enabled = true require_download_approval = true allowed_domains = ["skills.sh"] blocked_domains = ["internal.example"] install_root = "~/managed-skills" auto_expose_installed = true ``` Use this when: * one team wants a reviewed install root instead of a runtime-owned default location * download sources should be pinned to an explicit domain set * installed skills should become runtime-visible automatically across operator restarts ## Common Failure Cases | What happened | What it usually means | What to do next | | ------------------------------------------------- | -------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | | `skills fetch` is rejected before download starts | download approval or domain policy is blocking it | run `loong skills policy get`, then retry with `--approve-download` or update policy deliberately | | a skill installed but does not show up as usable | install succeeded, but the runtime exposure gate is still closed | set `auto_expose_installed = true` through `skills policy set` | | you do not know where skills should live | the runtime default is acceptable until you need a reviewed location | set `skills.install_root` in config when you need a stable explicit path | | `agent-browser` still does not run after install | shell policy or the runtime binary is still missing | remove `agent-browser` from `shell_deny`, verify `agent-browser open example.com`, then rerun `loong doctor` | ## Continue Reading * Continue to [Browser And Web Boundaries](/use-loong/browser-and-web-boundaries) when the failure is really a private-host, allowed-domain, or browser runtime gate problem. * Continue to [Tool Surface](/use-loong/tool-surface) when you want the broader truthful-tool contract around enablement surfaces. * Continue to [Doctor And Health](/get-started/doctor-and-health) when you want the health-first repair path after enabling skills or external browser automation. * Go back to [Configuration Patterns](/use-loong/configuration-patterns) when you want the wider operator config map.