The answer
Tool calls are the wrong first number. One test script or one agent stuck in a loop can produce hundreds of calls, while ten people who each come back for a second session are real adoption. Pick a window, such as the last 28 days, remove internal and test traffic, and read users, sessions and returning users together. Keep calls without a known user as their own group rather than dropping them.
| What you see | Real adoption when | False alarm when |
|---|---|---|
| Many tool calls | They come from many distinct users and sessions. | One user, script or retry loop produces most of them. |
| New users | They come from outside your team and test accounts. | They are teammates, CI jobs or a demo account. |
| Returning users | People are active again in a later week. | The same automation runs on a schedule. |
| Several AI clients | Different people use Claude, ChatGPT, Cursor or others. | One person tries the server in every client once. |
| Completed jobs | Sessions end with a result the user kept. | Calls succeed but the job never finishes. |
In Flowlines, the MCP server analytics view shows distinct users, sessions and tool calls for a window, and MCP user analytics shows each user's AI client, tools and return visits. The MCP server KPIs guide turns those counts into activation, retention and calls per completed job.
How can I tell real users from test traffic?
- Tag your own team, CI jobs and demo accounts, and exclude them before reporting adoption.
- Look at calls per user: one identity with thousands of calls and no other activity is usually automation.
- Check the AI client: a client name you have never heard of, or an empty one, often belongs to a script.
- Look at timing: calls at exact intervals around the clock are scheduled jobs, not people.
- Open a few sessions: a person asks varied questions; a script repeats the same call.
What if my MCP server shows no usage at all?
Check telemetry before concluding that nobody uses it. Make a known test call and confirm it arrives. If it does, check whether users can reach the server at all: authentication failures and client configuration errors stop people before their first tool call. If calls arrive but only for one or two tools, read why agents ignore your MCP tools. The MCP analytics guide covers the broader measurement model.
Published