Plugin bundles

OpenClaw can install plugins from four external ecosystems: the vendor-neutral Agent Plugins standard, plus Codex, Claude, and Cursor. These are called bundles - content and metadata packs that OpenClaw maps into native features like skills, hooks, and MCP tools.

Why bundles exist

Many useful plugins are published in the Agent Plugins, Codex, Claude, or Cursor format. Instead of requiring authors to rewrite them as native OpenClaw plugins, OpenClaw detects these formats and maps their supported content into the native feature set. You can install an Agent Plugins package, a Claude command pack, or a Codex skill bundle and use it immediately.

Install a bundle

# Local directoryopenclaw plugins install ./my-bundle # Archiveopenclaw plugins install ./my-bundle.tgz # Claude marketplaceopenclaw plugins marketplace list <source>openclaw plugins install <plugin> --marketplace <source>

openclaw plugins listopenclaw plugins inspect <id>

openclaw gateway restart

What OpenClaw maps from bundles

Not every bundle feature runs in OpenClaw. Here is what works and what is detected but not wired.

Supported now

Skill content

• Bundle skill roots load as normal OpenClaw skill roots.
• Claude commands/, agents/, and output-styles/ roots are treated as additional skill roots.
• Cursor .cursor/commands/ roots are treated as additional skill roots.

Claude markdown command files and Cursor command markdown both work through the normal OpenClaw skill loader.

Hook packs

Bundle hook roots are collection directories. Put each hook's HOOK.md and handler.ts or handler.js in its own child directory, such as hooks/my-hook/, and declare hooks/ as the root. Declaring the hook's leaf directory directly does not load it.

Plugin inspection lists these hook packs separately from detected JSON automation. Claude hooks/hooks.json remains in the declared capabilities, but does not appear as a supported hook. A bundle containing both layouts keeps its OpenClaw hook packs. Inspection does not execute handlers or prove that the running Gateway loaded them.

Embedded OpenClaw settings

Claude settings.json is imported as default embedded OpenClaw settings when the bundle is enabled. OpenClaw sanitizes shell override keys before applying them:

• shellPath
• shellCommandPrefix

Embedded OpenClaw LSP

• Enabled Claude bundles can contribute LSP server config.
• OpenClaw loads .lsp.json plus any manifest-declared lspServers paths.
• Bundle LSP config is merged into the effective embedded OpenClaw LSP defaults.
• Only supported stdio-backed LSP servers are runnable; unsupported transports still show up in openclaw plugins inspect <id>.

Detected but not executed

These are recognized and shown in diagnostics, but OpenClaw does not run them:

• Claude hooks/hooks.json automation
• Cursor .cursor/agents, .cursor/hooks.json, .cursor/rules
• Codex .app.json metadata beyond capability reporting

MCP for embedded OpenClaw

• Enabled bundles can contribute MCP server config.
• OpenClaw merges bundle MCP config into the effective embedded OpenClaw settings as mcpServers.
• OpenClaw exposes supported bundle MCP tools during embedded OpenClaw agent turns by launching stdio servers or connecting to HTTP servers.
• The coding and messaging tool profiles include bundle MCP tools by default; use tools.deny: ["bundle-mcp"] to opt out for an agent or gateway.
• Project-local embedded agent settings still apply after bundle defaults, so workspace settings can override bundle MCP entries when needed.
• Bundle MCP tool catalogs are sorted deterministically before registration, so upstream listTools() order changes do not thrash prompt-cache tool blocks.

Transports

MCP servers can use stdio or HTTP transport.

Stdio launches a child process:

{  "mcp": {    "servers": {      "my-server": {        "command": "node",        "args": ["server.js"],        "env": { "PORT": "3000" }      }    }  }}

HTTP connects to a running MCP server, defaulting to sse unless streamable-http is requested:

{  "mcp": {    "servers": {      "my-server": {        "url": "http://localhost:3100/mcp",        "transport": "streamable-http",        "headers": {          "Authorization": "Bearer ${MY_SECRET_TOKEN}"        },        "connectionTimeoutMs": 30000      }    }  }}

• transport accepts "streamable-http" or "sse"; omitted defaults to sse.
• type: "http" is a CLI-native downstream shape; use transport: "streamable-http" in OpenClaw config. openclaw mcp set and openclaw doctor --fix normalize the common alias.
• Only http: and https: URL schemes are allowed.
• headers values support ${ENV_VAR} interpolation.
• A server entry with both command and url is rejected.
• URL credentials (userinfo and query params) are redacted from tool descriptions and logs.
• connectionTimeoutMs overrides the default 30-second connection timeout for both stdio and HTTP transports. Request timeout defaults to 60 seconds and can be overridden with requestTimeoutMs.

Tool naming

OpenClaw registers bundle MCP tools with provider-safe names in the form serverName__toolName. For example, a server keyed "vigil-harbor" exposing a memory_search tool registers as vigil-harbor__memory_search.

• Characters outside A-Za-z0-9_- are replaced with -.
• Fragments that would start with a non-letter get a letter prefix, so numeric server keys such as 12306 become provider-safe tool prefixes.
• Server prefixes are capped at 30 characters.
• Full tool names are capped at 64 characters.
• Empty server names fall back to mcp.
• Colliding sanitized names are disambiguated with numeric suffixes.
• Final exposed tool order is deterministic by safe name, keeping repeated embedded-agent turns cache-stable.
• Profile filtering treats every tool from one bundle MCP server as plugin-owned by bundle-mcp, so profile allow/deny lists can reference either individual exposed tool names or the bundle-mcp plugin key.

Bundle formats

Detection precedence

OpenClaw checks for native plugin format first:

1. openclaw.plugin.json or a valid package.json with openclaw.extensions - treated as a native plugin
2. Client-specific bundle markers (.codex-plugin/, .cursor-plugin/, .claude-plugin/) - treated as a bundle in that format
3. A root plugin.json - treated as an Agent Plugins bundle
4. Default manifestless Claude layout (skills/, commands/, .mcp.json, ...) - treated as a Claude bundle

If a package carries both a client-specific marker and a root plugin.json, the client-specific format wins so its richer mappings (commands, hooks, settings) are preserved. If a directory contains both a native manifest and bundle markers, OpenClaw uses the native path. This prevents dual-format packages from being partially installed as bundles.

Runtime dependencies and cleanup

• Third-party compatible bundles do not get startup npm install repair. They should be installed through openclaw plugins install and ship everything they need in the installed plugin directory.
• OpenClaw-owned bundled plugins are either shipped lightweight in core or downloadable through the plugin installer. Gateway startup never runs a package manager for them.
• openclaw doctor --fix removes stale local bundled-plugin install records and can recover downloadable plugins that are missing from the local plugin index when config still references them.

Security

Bundles have a narrower trust boundary than native plugins:

• OpenClaw does not load arbitrary bundle runtime modules in-process.
• Skills and hook-pack paths must stay inside the plugin root (boundary-checked).
• Settings files are read with the same boundary checks.
• Supported stdio MCP servers may be launched as subprocesses.

This makes bundles safer by default, but you should still treat third-party bundles as trusted content for the features they do expose.

Troubleshooting

Related

• Install and Configure Plugins
• Building Plugins - create a native plugin
• Plugin Manifest - native manifest schema