Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools, data sources, and services. It is the "glue" that lets AI models like Claude or Copilot securely call external APIs, read files, query databases, and use custom tools — without bespoke integrations for every combination.

Why MCP Matters

Before MCP, connecting an AI to a tool meant writing a custom integration for every AI model × every tool combination. MCP standardizes this: write one MCP server per tool, and any MCP-compatible AI agent can use it.

This is directly analogous to what USB did for hardware peripherals — one standard port, many devices.

Core Concepts

MCP Server

A process that exposes tools, resources, and prompts over the MCP protocol. Examples:

• A filesystem server exposing read_file, write_file
• A GitHub server exposing create_issue, search_code
• A Tavily server exposing search, extract

MCP Client

The AI agent (or the runtime hosting the agent) that calls MCP server tools. In GitHub Agentic Workflows, the gh-aw runtime acts as the MCP client.

Transport Types

• stdio — local process with stdin/stdout (fastest, most common for local tools)
• HTTP/SSE — remote HTTP endpoint (cloud APIs, shared infrastructure)
• Docker containers — containerized MCP servers with env vars and volume mounts

MCP in GitHub Agentic Workflows

GitHub Agentic Workflows uses MCP as its primary tool integration mechanism. Every tool the agent can call — from GitHub operations to external APIs — is exposed via MCP.

Adding Custom MCP Servers

Custom MCP servers are declared in the workflow frontmatter under mcp-servers::

mcp-servers:
  notion:
    container: "mcp/notion"
    env:
      NOTION_TOKEN: "${{ secrets.NOTION_TOKEN }}"
    allowed:
      - "search_pages"
      - "get_page"
  deepwiki:
    url: "https://mcp.deepwiki.com/sse"
    allowed: ["*"]

Adding Servers from the Registry

# Browse available servers
gh aw mcp add

# Add a specific server
gh aw mcp add my-workflow makenotion/notion-mcp-server

Tool Filtering with allowed:

For security, restrict which tools an MCP server exposes to the agent:

mcp-servers:
  postgres:
    container: "mcp/postgres"
    allowed: ["query"]  # read-only; blocks writes

Use ["*"] to allow all tools (only for trusted, well-understood servers).

Official Reference Servers

Server	What It Does
Filesystem	Secure file operations with configurable access
Git	Read, search, and manipulate Git repositories
Fetch	Web content fetching and conversion for LLMs
Memory	Knowledge graph-based persistent memory
Sequential Thinking	Dynamic problem-solving through thought sequences
Time	Time and timezone conversion

A full registry of community MCP servers is at registry.modelcontextprotocol.io ↗.

Pre-configured MCP Servers in gh-aw

The gh-aw repository ships with ready-to-import MCP configurations in .github/workflows/shared/mcp/:

• Tavily — AI-native web search and extraction
• Jupyter — Execute code and manage notebooks
• Notion — Read/write Notion pages and databases
• Slack — Channel management and messaging
• Sentry — Error tracking and issue analysis
• DeepWiki — GitHub repository documentation search
• Microsoft Docs — Search official Microsoft documentation
• Azure, DataDog, Brave Search, MarkItDown, and more

MCP SDKs

Language	SDK Repository
TypeScript	modelcontextprotocol/typescript-sdk
Python	modelcontextprotocol/python-sdk
Go	modelcontextprotocol/go-sdk
Java	modelcontextprotocol/java-sdk
Rust	modelcontextprotocol/rust-sdk
C#	modelcontextprotocol/csharp-sdk
Kotlin, PHP, Ruby, Swift	Official SDKs available

MCP and AgentPages

AgentPages uses MCP via gh-aw to access:

• Tavily MCP — web search and URL extraction for research
• GitHub MCP (built-in) — creating PRs, committing files, reading repo contents
• Safe Outputs MCP — the sanitized write layer for GitHub operations

When you extend AgentPages with custom MCP servers, declare them in .github/workflows/research.md and recompile with gh aw compile.