Files
openclaw/docs/plugins/bundles.md
Peter Steinberger f7d7148cf0 docs: rewrite published docs grounded in current source (#100142)
Source-grounded rewrite of 529 published docs pages with per-unit information-loss verification: 1,713 factual corrections cited to src/**, generated surfaces regenerated, frontmatter titles preserved for i18n, release notes pages untouched. All docs gates green.

Closes #100141
2026-07-05 00:32:47 -04:00

314 lines
12 KiB
Markdown

---
summary: "Install and use Codex, Claude, and Cursor bundles as OpenClaw plugins"
read_when:
- You want to install a Codex, Claude, or Cursor-compatible bundle
- You need to understand how OpenClaw maps bundle content into native features
- You are debugging bundle detection or missing capabilities
title: "Plugin bundles"
---
OpenClaw can install plugins from three external ecosystems: **Codex**, **Claude**,
and **Cursor**. These are called **bundles** - content and metadata packs that
OpenClaw maps into native features like skills, hooks, and MCP tools.
<Info>
Bundles are **not** the same as native OpenClaw plugins. Native plugins run
in-process and can register any capability. Bundles are content packs with
selective feature mapping and a narrower trust boundary.
</Info>
## Why bundles exist
Many useful plugins are published in 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 a Claude command pack or a Codex skill bundle and use it
immediately.
## Install a bundle
<Steps>
<Step title="Install from a directory, archive, or marketplace">
```bash
# Local directory
openclaw plugins install ./my-bundle
# Archive
openclaw plugins install ./my-bundle.tgz
# Claude marketplace
openclaw plugins marketplace list <source>
openclaw plugins install <plugin> --marketplace <source>
```
`<source>` is a local marketplace path/repo or a git/GitHub source.
</Step>
<Step title="Verify detection">
```bash
openclaw plugins list
openclaw plugins inspect <id>
```
Bundles show `Format: bundle` plus a `Bundle format:` value of `codex`,
`claude`, or `cursor`.
</Step>
<Step title="Restart and use">
```bash
openclaw gateway restart
```
Mapped features (skills, hooks, MCP tools, LSP defaults) are available in the next session.
</Step>
</Steps>
## What OpenClaw maps from bundles
Not every bundle feature runs in OpenClaw today. Here is what works and what
is detected but not yet wired.
### Supported now
| Feature | How it maps | Applies to |
| ------------- | ------------------------------------------------------------------------------------------------- | -------------- |
| Skill content | Bundle skill roots load as normal OpenClaw skills | All formats |
| Commands | `commands/` and `.cursor/commands/` treated as skill roots | Claude, Cursor |
| Hook packs | OpenClaw-style `HOOK.md` + `handler.ts` layouts | Codex |
| MCP tools | Bundle MCP config merged into embedded OpenClaw settings; supported stdio and HTTP servers loaded | All formats |
| LSP servers | Claude `.lsp.json` and manifest-declared `lspServers` merged into embedded OpenClaw LSP defaults | Claude |
| Settings | Claude `settings.json` imported as embedded OpenClaw defaults | Claude |
#### Skill content
- Bundle skill roots load as normal OpenClaw skill roots.
- Claude `commands/` 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 work **only** when they use the normal OpenClaw hook-pack
layout: `HOOK.md` plus `handler.ts` or `handler.js`. Today this is primarily
the Codex-compatible case.
#### 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:
```json
{
"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:
```json
{
"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.
#### 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 today; 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 `agents`, `hooks/hooks.json` automation, `outputStyles`
- Cursor `.cursor/agents`, `.cursor/hooks.json`, `.cursor/rules`
- Codex `.app.json` metadata beyond capability reporting
## Bundle formats
<AccordionGroup>
<Accordion title="Codex bundles">
Markers: `.codex-plugin/plugin.json`
Optional content: `skills/`, `hooks/`, `.mcp.json`, `.app.json`
Codex bundles fit OpenClaw best when they use skill roots and OpenClaw-style
hook-pack directories (`HOOK.md` + `handler.ts`).
</Accordion>
<Accordion title="Claude bundles">
Two detection modes:
- **Manifest-based:** `.claude-plugin/plugin.json`
- **Manifestless:** default Claude layout (`skills/`, `commands/`, `agents/`, `hooks/`, `.mcp.json`, `.lsp.json`, `settings.json`)
Claude-specific behavior:
- `commands/` is treated as skill content
- `settings.json` is imported into embedded OpenClaw settings (shell override keys are sanitized)
- `.mcp.json` exposes supported stdio tools to embedded OpenClaw
- `.lsp.json` plus manifest-declared `lspServers` paths load into embedded OpenClaw LSP defaults
- `hooks/hooks.json` is detected but not executed
- Custom component paths in the manifest are additive; they extend defaults, not replace them
</Accordion>
<Accordion title="Cursor bundles">
Markers: `.cursor-plugin/plugin.json`
Optional content: `skills/`, `.cursor/commands/`, `.cursor/agents/`, `.cursor/rules/`, `.cursor/hooks.json`, `.mcp.json`
- `.cursor/commands/` is treated as skill content
- `.cursor/rules/`, `.cursor/agents/`, and `.cursor/hooks.json` are detect-only
</Accordion>
</AccordionGroup>
## 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. Bundle markers (`.codex-plugin/`, `.claude-plugin/`, or default Claude/Cursor layout) - treated as a **bundle**
If a directory contains both, 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
<AccordionGroup>
<Accordion title="Bundle is detected but capabilities do not run">
Run `openclaw plugins inspect <id>`. If a capability is listed but marked as
not wired, that is a product limit, not a broken install.
</Accordion>
<Accordion title="Claude command files do not appear">
Make sure the bundle is enabled and the markdown files are inside a detected
`commands/` or `skills/` root.
</Accordion>
<Accordion title="Claude settings do not apply">
Only embedded OpenClaw settings from `settings.json` are supported. OpenClaw does
not treat bundle settings as raw config patches.
</Accordion>
<Accordion title="Claude hooks do not execute">
`hooks/hooks.json` is detect-only. If you need runnable hooks, use the
OpenClaw hook-pack layout or ship a native plugin.
</Accordion>
</AccordionGroup>
## Related
- [Install and Configure Plugins](/tools/plugin)
- [Building Plugins](/plugins/building-plugins) - create a native plugin
- [Plugin Manifest](/plugins/manifest) - native manifest schema