MCP Servers

What is MCP?

The Model Context Protocol (MCP) is an open standard that lets AI agents call external tools and access data sources. An MCP server is a lightweight process or service that exposes a set of tools – the agent discovers them automatically and can invoke them during conversations.

This is how Polyclaw gets capabilities like browser automation, Azure resource management, documentation search, and GitHub integration without any of that logic living inside the agent itself. You enable the MCP servers you need, and the agent gains their tools.

Managing MCP Servers

The web dashboard is the easiest way to discover, add, and manage MCP servers.

Discover and Browse

The MCP Servers page lists all configured servers along with their type, status, and a toggle to enable or disable each one.

MCP server list

Marketplace

The discovery page lets you browse a catalog of available MCP servers and enable them with one click.

MCP server marketplace

Adding a Server

You can add any MCP server – local or remote – through the dashboard. Specify whether it is a local subprocess or a remote HTTP/SSE endpoint, provide the connection details, and save.

Adding a remote or local MCP server

Slash Commands

/mcp                    # List all MCP servers
/mcp enable <id>        # Enable a server
/mcp disable <id>       # Disable a server
/mcp add <name> <url>   # Add a new server definition
/mcp remove <id>        # Remove a server

API

GET  /api/mcp/servers                # List servers
POST /api/mcp/servers                # Add server
PUT  /api/mcp/servers/<id>           # Toggle / update server
DELETE /api/mcp/servers/<id>         # Remove server

Using MCP Tools in Conversations

Once an MCP server is enabled, the agent can use its tools immediately – no extra prompting required. The agent discovers available tools at session start and decides when to call them based on your request.

Example: Microsoft Learn

Enable the Microsoft Learn MCP server, then ask the agent to look something up in the docs.

Using Microsoft Learn MCP in chat

Example: Playwright Browser

The Playwright MCP server gives the agent a full headless browser. It can navigate pages, fill forms, take screenshots, and extract content.

Using Playwright browser automation in chat

How It Works Under the Hood

On session creation, all enabled MCP servers are passed to the Copilot SDK
The SDK connects to each server and discovers its tools via tools/list
MCP server descriptions are included in the agent’s system prompt
During conversation, the LLM can call any tool exposed by connected servers
Tool results are returned to the LLM as context for generating responses

Server Types

Polyclaw supports four transport types:

Type	Description
`local`	Subprocess spawned by Polyclaw, communicates via stdio
`stdio`	External process with stdin/stdout transport
`http`	Remote HTTP endpoint
`sse`	Server-Sent Events endpoint

Built-in MCP Servers

Polyclaw ships with several pre-configured MCP servers:

Server	Type	Default	Purpose
Playwright	local	Enabled	Browser automation (web search, scraping, UI testing)
Microsoft Learn	http	Disabled	Search Microsoft documentation
Azure MCP Server	local	Disabled	Azure resource management (requires `az login`)
GitHub MCP Server	local	Disabled	GitHub API integration (requires `gh auth login`)

Configuration Storage

MCP server definitions are stored in ~/.polyclaw/mcp_servers.json. Each entry specifies a server type, connection details, and metadata. You rarely need to edit this file directly – the dashboard and slash commands manage it for you.

Adding Custom MCP Servers

Any process that implements the MCP protocol can be added. The server must:

Accept connections via the configured transport (stdio, HTTP, or SSE)
Respond to tools/list to enumerate available tools
Handle tools/call to execute tool invocations
Return structured results

See the MCP specification for the full protocol definition.