MCP in VS Code: how to connect GitHub Copilot to external tools

VS Code added Model Context Protocol support in GitHub Copilot's agent mode starting in version 1.99. Once you wire up an MCP server, Copilot can call external tools — searching documentation, querying APIs, checking indexed knowledge — directly inside a Copilot Chat conversation. No extension to install, no additional subscription.

This guide shows how to connect AgentReady so that GitHub Copilot can query any indexed website from inside VS Code — useful for library docs, internal tools, or any site that isn't in Copilot's training data.

What VS Code MCP actually does

MCP support in VS Code means GitHub Copilot agent mode can call MCP tools during a conversation. When you ask Copilot something that requires external data — “what does the Supabase docs say about realtime subscriptions?” — Copilot can call an MCP tool that retrieves the answer from indexed content rather than falling back to (potentially stale) training data.

The tools are visible to Copilot automatically once the MCP server is registered. You don't need to invoke them explicitly — Copilot decides when to call them based on what your question needs.

Step 1 — Check your VS Code version

MCP support requires VS Code 1.99 or later and the GitHub Copilot extension. Check with Help → About in the menu bar. Update via Help → Check for Updates if needed.

MCP tools are only available in Copilot Chat agent mode. Make sure you're using agent mode (@workspace or the agent toggle in the chat panel) rather than inline completions.

Step 2 — Add the MCP server config

You can configure MCP servers at two scopes — workspace (project-level) or user (global):

Workspace scope — create .vscode/mcp.json in your project root:

{
  "servers": {
    "agentready": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@agentreadyweb/mcp"]
    }
  }
}

Commit this file so every developer on the project gets AgentReady's tools in their Copilot sessions automatically.

User scope — open VS Code settings (Cmd+,), search for MCP, and click Edit in settings.json. Add under the mcp.servers key:

"mcp.servers": {
  "agentready": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@agentreadyweb/mcp"]
  }
}

Step 3 — Verify the server is connected

Open Copilot Chat (Ctrl+Alt+I / Cmd+Shift+I) and switch to agent mode. You should see AgentReady's tools listed in the tools panel (the wrench icon in the chat input bar). The four tools are:

• list_sites — shows all indexed sites
• ask_site — queries a site with a natural-language question
• submit_site — indexes a new site for the first time
• refresh_site — re-crawls a site to pick up new content

If you don't see them: open the Command Palette (Cmd+Shift+P), run MCP: List Servers, and check the AgentReady entry shows as running.

Step 4 — Query docs in Copilot Chat

With the server running, ask Copilot questions that reference external docs:

# Copilot Chat (agent mode)
What does the Tailwind docs say about container queries?
Check the Supabase docs — how do I set up Row Level Security for a multi-tenant app?
How does Stripe handle idempotency keys according to their docs?
What's the Next.js App Router way to handle dynamic routes with search params?

Copilot will call ask_sitewith the inferred domain and question. You'll see the tool call in the chat alongside the response, including which pages were cited.

If you ask about a site that isn't indexed yet, ask Copilot to submit it first: “use AgentReady to index prisma.io/docs.” The first query after indexing takes about 60 seconds.

VS Code MCP vs Cursor MCP — what's different

The config format and file locations differ, but the runtime behaviour is the same — both run the MCP server as a local subprocess and pass tool calls over stdio. A few practical differences:

• Discovery: Cursor scans ~/.cursor/mcp.json; VS Code scans .vscode/mcp.json and the user settings file. Both support workspace-level config, which is useful for team sharing.
• Agent mode requirement: VS Code only runs MCP tools in Copilot agent mode. Cursor uses MCP in all its AI modes including inline chat and composer.
• Tool approval: VS Code prompts for confirmation before running each tool call by default. You can disable this per-server in the MCP settings if it gets noisy during heavy usage.

Sharing with your team

The workspace config approach (.vscode/mcp.json) is the right way to share MCP tools across a team. Add it to your dotfiles or project repo and every developer gets the tools without any individual setup. The @agentreadyweb/mcp package is pulled via npx at startup — no global installation required.

Sites indexed by AgentReady are shared across all clients. If one team member submits a site, it's immediately queryable by everyone using the same MCP server — whether they're in VS Code, Cursor, Claude Desktop, or Claude Code.