Agents: run bundle MCP tools in embedded Pi (#48611)

* Agents: run bundle MCP tools in embedded Pi

* Plugins: fix bundle MCP path resolution

* Plugins: warn on unsupported bundle MCP transports

* Commands: add embedded Pi MCP management

* Config: move MCP management to top-level config
This commit is contained in:
Vincent Koc
2026-03-16 21:46:05 -07:00
committed by GitHub
parent 38bc364aed
commit 06459ca0df
37 changed files with 2051 additions and 30 deletions

View File

@@ -104,11 +104,15 @@ loader. Cursor command markdown works through the same path.
- `HOOK.md`
- `handler.ts` or `handler.js`
#### MCP for CLI backends
#### MCP for Pi
- enabled bundles can contribute MCP server config
- current runtime wiring is used by the `claude-cli` backend
- OpenClaw merges bundle MCP config into the backend `--mcp-config` file
- OpenClaw merges bundle MCP config into the effective embedded Pi settings as
`mcpServers`
- OpenClaw also exposes supported bundle MCP tools during embedded Pi agent
turns by launching supported stdio MCP servers as subprocesses
- project-local Pi settings still apply after bundle defaults, so workspace
settings can override bundle MCP entries when needed
#### Embedded Pi settings
@@ -133,7 +137,6 @@ diagnostics/info output, but OpenClaw does not run them yet:
- Cursor `.cursor/agents`
- Cursor `.cursor/hooks.json`
- Cursor `.cursor/rules`
- Cursor `mcpServers` outside the current mapped runtime paths
- Codex inline/app metadata beyond capability reporting
## Capability reporting
@@ -153,7 +156,8 @@ Current exceptions:
- Claude `commands` is considered supported because it maps to skills
- Claude `settings` is considered supported because it maps to embedded Pi settings
- Cursor `commands` is considered supported because it maps to skills
- bundle MCP is considered supported where OpenClaw actually imports it
- bundle MCP is considered supported because it maps into embedded Pi settings
and exposes supported stdio tools to embedded Pi
- Codex `hooks` is considered supported only for OpenClaw hook-pack layouts
## Format differences
@@ -195,6 +199,8 @@ Claude-specific notes:
- `commands/` is treated like skill content
- `settings.json` is imported into embedded Pi settings
- `.mcp.json` and manifest `mcpServers` can expose supported stdio tools to
embedded Pi
- `hooks/hooks.json` is detected, but not executed as Claude automation
### Cursor
@@ -246,7 +252,9 @@ Current behavior:
- bundle discovery reads files inside the plugin root with boundary checks
- skills and hook-pack paths must stay inside the plugin root
- bundle settings files are read with the same boundary checks
- OpenClaw does not execute arbitrary bundle runtime code in-process
- supported stdio bundle MCP servers may be launched as subprocesses for
embedded Pi tool calls
- OpenClaw does not load arbitrary bundle runtime modules in-process
This makes bundle support safer by default than native plugin modules, but you
should still treat third-party bundles as trusted content for the features they