MCP server / Connection guide

Connect Serge to your
LLM client

This guide connects Claude Desktop, Claude on the web, or Cursor to your Serge agent-analytics workspace over MCP. Choose an auth method, configure the client, verify the connection, and ask your first question.

Need

A Serge workspace with at least one registered site, and an LLM client that speaks MCP (Claude Desktop, Claude web, or Cursor).

Endpoint

All clients connect to https://mcp.serge.ai.

Best for

Asking one-off questions about your agent traffic without clicking through dashboard filters.

1

Choose how to authenticate

Pick one of two methods. You don't need both.

API key

Mint a key in Settings → API keys (it starts with sk_serge_) with the scans:read and traffic:read scopes. Best for Claude Desktop and Cursor, where the key lives in a config file.

OAuth

Add https://mcp.serge.ai as a connector in Claude and sign in with your Serge account. Best when you'd rather log in than manage a key, or you're on Claude web.

2

Configure your client

Open your Claude Desktop config file (Cursor uses the same JSON shape under its MCP settings):

macOS ~/Library/Application Support/Claude/claude_desktop_config.json
Windows %APPDATA%\Claude\claude_desktop_config.json

Add Serge under the mcpServers object. Replace the placeholder with your sk_serge_ key:

Using OAuth instead? Skip the config file: in Claude, open connector settings, add https://mcp.serge.ai, and complete the sign-in.

3

Restart and verify the connection

Restart your client so it loads the new server. The Serge tools should appear in the tool picker.

Confirm the connection is bound to the right workspace by asking your client to call whoami:

It returns your workspace and the scopes on the key. If the scopes are missing, mint a new key with scans:read and traffic:read.

4

Ask your first question

Use plain language. Your client maps the question to the right tool and fills in the period and domain.

If a tool replies site_not_found, ask your client to call list_sites and use a domain it returns.

What you can ask

Traffic"How many AI agents visited yourstore.com this week, and which platforms?"
Buy-intent"What share of agent traffic on yourstore.com was real buy-intent versus crawlers?"
Failures"Show me the sessions where agents abandoned checkout on yourstore.com."
Scan"What's the latest scan score for yourstore.com, and what should I fix first?"

More example prompts

Which entry pages do AI agents land on most on yourstore.com?

Did any platform cryptographically verify its identity on yourstore.com last week?

Walk me through one failing session on yourstore.com page by page.

Compare buy-intent traffic on yourstore.com over the last 7 days versus 30 days.

Turn findings into fixes

Once a session journey or scan surfaces a blocker, paste it into Claude, Cursor, or your coding agent with a prompt like this:

Next steps