Plugins★

In LLM systems, a plugin is a modular package that bundles multiple extension capabilities together so an agent can immediately execute domain-specific tasks. Rather than writing custom API integration code for every new utility, a developer installs a plugin to load tools, rules, and prompt instructions at runtime.

Jargon Decoder: MCP vs. Skill vs. Agent vs. Plugin

AI engineering has several overlapping terms. Here is how they break down in practice:

1. MCP (Model Context Protocol):
   • What it is: A transport standard protocol (JSON-RPC over stdin/stdout or SSE).
   • Role: It is the hardware connector or "USB cable" between the client (IDE/Agent) and the server (database/API).
2. Skill:
   • What it is: A specific behavioral set of instructions (markdown guidelines or helper scripts).
   • Role: It represents the "recipes" or "capabilities" the agent learns to complete a specific task (e.g. running a genomic variant search).
3. Agent:
   • What it is: The autonomous loop orchestrator.
   • Role: The "brain" that accepts prompts, plans actions, decides which tool to call, and processes tool outputs statefully.
4. Plugin:
   • What it is: A packaged bundle of all the above.
   • Role: The delivery container. A single plugin can bundle multiple skills, install executable tool schemas (which communicate via MCP), and configure subagent prompt definitions for a specific ecosystem.

Technical Architecture Comparison

Consider an AI development workspace where we load a code-quality assistant:

• The Client (Agent): The master LLM loop managing file writes.
• The Plugin (chrome-devtools-plugin): The installer bundle containing devtools integrations.
• The MCP Server (chrome-devtools-mcp): The server interface exposing the page navigation and screenshot tools.
• The Skill (alphafold-fetch): The specialized markdown rulebook directing the agent on how to interpret protein coordinate files.

  ┌────────────────────────────────────────────────────────┐
  │                   PLUGIN BUNDLE                        │
  │  ┌──────────────┐  ┌─────────────────┐  ┌───────────┐  │
  │  │  Skills /    │  │   MCP Server    │  │ Subagent  │  │
  │  │  Instructions│  │   (Tool Schemas)│  │ Prompt    │  │
  │  └──────────────┘  └────────┬────────┘  └───────────┘  │
  └─────────────────────────────┼──────────────────────────┘
                                │ (JSON-RPC Protocol)
                                ▼
                       [ AGENT RUNTIME ]

Configuration in Code

When loading plugins into an agent frame, developers register them at startup so their endpoints are exposed to the reasoning loop:

`typescript import { Agent } from 'antigravity-sdk'; import { ChromeDevToolsPlugin } from 'chrome-devtools-plugin';

const agent = new Agent({ model: 'claude-3-5-sonnet', plugins: [ new ChromeDevToolsPlugin({ headless: true }) // Automatically mounts MCP tools & skills ] }); `