Skip to main content
Plugins extend OpenClaw with channels, model providers, agent harnesses, tools, skills, speech, realtime transcription, voice, media understanding, generation, web fetch, web search, and other runtime capabilities. Use this page to install a plugin, restart the Gateway, verify the runtime loaded it, and route common setup failures. For command-only examples, see Manage plugins. For the generated inventory of bundled, official external, and source-only plugins, see Plugin inventory.

​
Requirements

  • an OpenClaw checkout or installation with the openclaw CLI available
  • network access to the selected source (ClawHub, npm, or a git host)
  • any plugin-specific credentials, config keys, or OS tools named by that plugin’s setup docs
  • permission for the Gateway that serves your channels to reload or restart

​
Quick start

1

Find the plugin

Search ClawHub for public plugin packages:
ClawHub is the primary discovery surface for community plugins. During the launch cutover, ordinary bare package specs still install from npm unless they match an official plugin id. Raw @openclaw/* specs that match a bundled plugin resolve to that bundled copy. Use an explicit source prefix when you need one source specifically.
2

Install the plugin

Treat plugin installs like running code. Prefer pinned versions for reproducible production installs. ClawHub packages and OpenClaw’s bundled/official catalog are trusted sources. New arbitrary npm, git, local path/archive, npm-pack:, or marketplace sources require --force in noninteractive installs after you review and trust the source.
3

Configure and enable it

Configure plugin-specific settings under plugins.entries.<id>.config. Enable the plugin if it is not already enabled:
If plugins.allow is set, the installed plugin id must be in that list before the plugin can load. openclaw plugins install adds the installed id to an existing plugins.allow list and removes the same id from plugins.deny so the explicit install can load after restart.
4

Let the Gateway reload

Installing, updating, or uninstalling plugin code requires a Gateway restart. A managed Gateway with config reload enabled detects the changed plugin install record and restarts automatically. Otherwise, restart it yourself:
Enable/disable update config and the cold registry. A runtime inspect is still the clearest proof of live runtime surfaces.
5

Verify runtime registration

Use --runtime to prove registered tools, hooks, services, Gateway methods, or plugin-owned CLI commands. Plain inspect is a cold manifest and registry check only.

​
Configuration

​
Choose an install source

Bare package specs have special compatibility behavior: a bare name that matches a bundled plugin id uses that bundled source; a bare name that matches an official external plugin id uses the official package catalog; any other bare spec installs through npm during the launch cutover. Raw @openclaw/* specs that match bundled plugins also resolve to the bundled copy before npm fallback. Use npm:@openclaw/<plugin>@<version> to deliberately install the external npm package instead of the bundled copy. Use clawhub:, npm:, git:, or npm-pack: for deterministic source selection. See openclaw plugins for the full command contract. For npm installs, unpinned specs and @latest choose the newest stable package that advertises compatibility with this OpenClaw build. If npm’s current latest release declares a newer openclaw.compat.pluginApi or openclaw.install.minHostVersion than this build supports, OpenClaw scans older stable versions and installs the newest one that fits. Exact versions and explicit channel tags such as @beta stay pinned to the selected package and fail when incompatible.

​
Operator install policy

Configure security.installPolicy to run a trusted local policy command before a plugin install or update proceeds. The policy receives metadata plus the staged source path and can allow or block the install. It covers both CLI and Gateway-backed install/update paths. Plugin before_install hooks run later, and only in OpenClaw processes where plugin hooks are loaded, so use security.installPolicy for operator-owned install decisions instead. The deprecated --dangerously-force-unsafe-install flag is accepted for compatibility but is a no-op: it does not bypass install policy or OpenClaw’s built-in plugin dependency denylist. See Skills config for the shared security.installPolicy exec schema used by both skills and plugins.

​
Configure plugin policy

The common plugin config shape is:
Key policy rules:
  • plugins.enabled: false disables all plugins and skips discovery/load work. Stale plugin references stay inert while this is active; re-enable plugins before running doctor cleanup if you want stale ids removed.
  • plugins.deny wins over allow and per-plugin enablement.
  • plugins.allow is an exclusive allowlist. Plugin-owned tools outside the allowlist stay unavailable even when tools.allow includes "*".
  • plugins.entries.<id>.enabled: false disables one plugin while keeping its config.
  • plugins.load.paths adds explicit local plugin files or directories. Managed plugins install local paths must be plugin directories or archives; use plugins.load.paths for standalone plugin files.
  • Workspace-origin plugins are disabled by default; explicitly enable or allowlist them before using local workspace code.
  • Bundled plugins follow their built-in default-on/default-off metadata unless config explicitly overrides it.
  • plugins.slots.<slot> (memory or contextEngine) picks one plugin for an exclusive category. Slot selection counts as explicit activation and force-enables the selected plugin for that slot, even if it would otherwise be opt-in. plugins.deny and plugins.entries.<id>.enabled: false still block it.
  • Bundled opt-in plugins can auto-activate when config names one of their owned surfaces, such as a provider/model ref, channel config, CLI backend, or agent harness runtime.
  • OpenAI-family Codex routing keeps provider and runtime plugin boundaries separate: legacy Codex model refs are legacy config that doctor repairs, while the bundled codex plugin owns Codex app-server runtime for canonical openai/* agent refs, explicit agentRuntime.id: "codex", and legacy codex/* refs.
When plugins.allow is unset and non-bundled plugins are auto-discovered from the workspace or global plugin roots, startup logs plugins.allow is empty; discovered non-bundled plugins may auto-load: ... with the discovered plugin ids and, for short lists, a minimal plugins.allow snippet. Run openclaw plugins list --enabled --verbose or openclaw plugins inspect <id> on the listed plugin id before copying trusted plugins into openclaw.json. The same trust-pinning applies when diagnostics say a plugin loaded without install/load-path provenance: inspect that plugin id, then pin it in plugins.allow or reinstall from a trusted source so OpenClaw records install provenance. Run openclaw doctor or openclaw doctor --fix when config validation reports stale plugin ids, allowlist/tool mismatches, or legacy bundled plugin paths.

​
Understand plugin formats

OpenClaw recognizes two plugin formats: Both formats appear in openclaw plugins list, openclaw plugins inspect, openclaw plugins enable, and openclaw plugins disable. See Plugin bundles for the bundle compatibility boundary and Building plugins for native plugin authoring.

​
Plugin hooks

Plugins can register hooks at runtime through two different APIs:
  • api.on(...) typed hooks for runtime lifecycle events. This is the preferred surface for middleware, policy, message rewriting, prompt shaping, and tool control.
  • api.registerHook(...) for the internal hook system described in Hooks. This is mainly for coarse command/lifecycle side effects and compatibility with existing HOOK-style automation.
Quick rule: if the handler needs priority, merge semantics, or block/cancel behavior, use typed hooks. If it just reacts to command:new, command:reset, message:sent, or similar coarse events, api.registerHook is fine. Plugin-managed internal hooks show up in openclaw hooks list with plugin:<id>. You cannot enable or disable them through openclaw hooks; enable or disable the plugin instead.

​
Verify the active Gateway

openclaw plugins list and plain openclaw plugins inspect read cold config, manifest, and registry state. They do not prove that an already-running Gateway has imported the same plugin code. When a plugin appears installed but live chat traffic does not use it:
Managed Gateways restart automatically after plugin install, update, and uninstall changes that alter plugin source. On VPS or container installs, make sure any manual restart targets the actual openclaw gateway run child that serves your channels, not only a wrapper or supervisor.

​
Troubleshooting

When stale plugin config still names a no-longer-discoverable channel plugin, config validation downgrades that channel key to a warning instead of a hard failure, so Gateway startup can still serve every other channel. Run openclaw doctor --fix to remove stale plugin and channel entries. Unknown channel keys without stale-plugin evidence still fail validation so typos stay visible. For intentional channel replacement, the preferred plugin should declare channelConfigs.<channel-id>.preferOver with the legacy or lower-priority plugin id. If both plugins are explicitly enabled, OpenClaw keeps that request and reports duplicate channel/tool diagnostics instead of silently choosing one owner. If an installed package reports that it requires compiled runtime output for TypeScript entry ..., the package was published without the JavaScript files OpenClaw needs at runtime. Update or reinstall after the publisher ships compiled JavaScript, or disable/uninstall the plugin until then.

​
Blocked plugin path ownership

If diagnostics say blocked plugin candidate: suspicious ownership (... uid=1000, expected uid=0 or root) and validation follows with plugin present but blocked, OpenClaw found plugin files owned by a different Unix user than the process loading them. Keep the plugin config in place; fix the filesystem ownership or run OpenClaw as the same user that owns the state directory. For Docker installs, the official image runs as node (uid 1000), so the host bind-mounted OpenClaw config and workspace directories should normally be owned by uid 1000:
If you intentionally run OpenClaw as root, repair the managed plugin root to root ownership instead:
After fixing ownership, rerun openclaw doctor --fix or openclaw plugins registry --refresh so the persisted plugin registry matches the repaired files.

​
Slow plugin tool setup

If agent turns appear to stall while preparing tools, enable trace logging and check for plugin tool factory timing lines:
Look for:
The summary lists total factory time and the slowest plugin tool factories, including plugin id, declared tool names, result shape, and whether the tool is optional. Slow lines are promoted to warnings when a single factory takes at least 1s or total plugin tool factory prep takes at least 5s. OpenClaw caches successful plugin tool factory results for repeated resolutions with the same effective request context. The cache key includes the effective runtime config, workspace and agent id, sandbox policy, browser settings, delivery context, requester identity, and ownership state, so factories that depend on those trusted fields re-run when the context changes. If timings stay high, the plugin may be doing expensive work before returning its tool definitions. If one plugin dominates the timing, inspect its runtime registrations:
Then update, reinstall, or disable that plugin. Plugin authors should move expensive dependency loading behind the tool execution path instead of doing it inside the tool factory. For dependency roots, package metadata validation, registry records, startup reload behavior, and legacy cleanup, see Plugin dependency resolution.