Plugins & MCP

Polyclaw extends its capabilities through two complementary systems: Plugins and MCP Servers.

What is MCP?

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools and data sources. MCP servers expose callable functions that the LLM can invoke during conversations.

Polyclaw supports four MCP server types:

Type	Description
`local`	Spawned as a subprocess, communicates via stdio
`stdio`	External process with stdin/stdout transport
`http`	Remote HTTP endpoint
`sse`	Server-Sent Events endpoint

What are Plugins?

Plugins are bundles that package skills, metadata, and dependency declarations together. They provide a higher-level abstraction for distributing and managing sets of related capabilities.

Each plugin contains:

A PLUGIN.json manifest
One or more skill directories
Optional dependency declarations (CLI tools, pip packages)
Optional setup flow (a skill that guides first-time configuration)

How They Work Together

Plugin
  |-- PLUGIN.json (metadata, dependencies)
  |-- skills/
  |     |-- skill-a/SKILL.md
  |     |-- skill-b/SKILL.md

Plugins and MCP servers are managed independently but work well together. For example, a plugin can ship a skill that instructs the agent to use Playwright for browser automation, or a skill that relies on the WorkIQ MCP server for Microsoft 365 data.

When a plugin is enabled:

Its skill directories are copied to the user skills directory
Dependencies are checked

When disabled:

Skill directories are removed

Sections

MCP Servers – Configuration and built-in MCP servers
Creating Plugins – Plugin manifest format and development guide
Bundled Plugins – Documentation for included plugins

MCP Servers → Creating Plugins → Bundled Plugins →