MCP Hands-on: Install, Call Tools, See Real Results

When finished you should see all three (missing any = not connected):

1. Connected in settings: Settings → MCP shows green Connected (or Tools > 0)
2. Tool calls in chat: e.g. list_directory / read_file
3. Real local data in the reply—not invented paths

We demo Filesystem MCP. On any AgentHub MCP page, copy that page’s Cursor config—same workflow.

Goal

Let Cursor Agent list and read files in a folder you allow:

• “List files under ~/Documents/demo-mcp”
• “Read the first 30 lines of README.md there”

Need

• Cursor (recent stable)
• Node.js (node -v works; most MCPs start via npx)
• A practice folder (don’t mount your whole disk)

mkdir -p ~/Documents/demo-mcp
echo "# MCP Demo" > ~/Documents/demo-mcp/README.md
echo "hello from mcp" > ~/Documents/demo-mcp/notes.txt

Find on AgentHub

Open /mcp, search filesystem, open a detail page. Prefer the JSON from that page’s Cursor tab.

1. Open the MCP detail page (filesystem for this demo)
2. Install section → Cursor tab → Copy

Illustrative JSON:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/YOU/Documents/demo-mcp"
      ]
    }
  }
}

How to read this config

Same shape whether the server is filesystem or sbuilder-mcp as in the diagram:

Notes

• Last args entry is often the allowed root—set to your demo-mcp absolute path
• Fix env placeholders before saving or the server won’t start

Project-scoped (recommended)

In the open project root:

mkdir -p .cursor

Create/edit .cursor/mcp.json and merge mcpServers (don’t wipe existing keys).

Minimal example:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/YOU/Documents/demo-mcp"
      ]
    }
  }
}

Then:

1. Reload Window
2. Settings → MCP
3. Confirm Connected (first npx may take a while)

UI you should see

• Entry named filesystem (or your key)
• Tools: list_directory, read_file, …
• Red/Error → check logs: bad path, Node not on PATH, invalid JSON

In Agent/Composer chat (tools enabled):

Prompt A｜list

Use filesystem MCP list_directory on ~/Documents/demo-mcp. Report only tool results—don’t invent files.

Expected A

• Tool-call chip with that path
• Both README.md and notes.txt
• No tool call → MCP not connected / tools disabled

Prompt B｜read

Read notes.txt via MCP and quote the content.

Expected B: read_file call + reply contains hello from mcp

Prompt C｜contrast

Disable MCP → ask again (can’t see folder) → re-enable → real content returns. That’s MCP.

Any MCP on AgentHub: copy → merge + secrets → reload/Connected → verify with a tool-forcing prompt.

Verification by type

Detail sidebars link back here for the same checklist.

Claude Code

Copy the Claude Code command from AgentHub, run it, then verify with the same list/read prompts. Use claude mcp list.

Other IDEs

Use that client’s tab on the detail page. Success criterion: tool calls + real data.

Shorter notes: Cursor MCP setup, Claude Code MCP setup.

After this lab, every MCP detail page uses the same accept criteria: tool call + real result.