What MCP activity can tell us about adoption

In one sampled week, a reported coding client generated well over 100,000 successful inference requests across dozens of organisations. One organisation accounted for most of the volume. The request total looked substantial, but it gave me little sense of how widely people were using the service.

That was a separate traffic sample from the MCP data I worked with. For MCP, I used a local prototype that grouped an anonymised snapshot by account. I wanted to see which accounts came back, what they tried and which operations succeeded. The prototype ran on prepared summaries in a development environment.

Associating activity with accounts

The host application can attach an authenticated organisation to each operation. That gives the provider an account to investigate even when the client name is ambiguous.

The reported client label has a different evidential basis. It records the name the client supplied, and a proxy or SDK may obscure the application behind it. Account authentication does not verify the claimed client vendor or identify the person operating it.

Grouping by account lets an engineer examine which capabilities it reached, whether activity recurred and which operations were refused. The prototype included an account with activity under multiple reported client labels. Its operations were model-list and context reads. This established activity across labels, while leaving broader feature adoption and the number of people or underlying applications unknown.

The snapshot held separate client and tool totals. Assigning particular tools to each client would require a combined rollup that the prototype did not have.

Interpreting successful operations

Read-only activity can fulfil a complete task. An agent asked to inspect usage or check model access may have no reason to change anything. The presence of a mutation therefore makes a poor general criterion for successful use.

In a sample of logs, I found hundreds of authenticated MCP tool attempts. Most operations returned an inner success response, while some were refused or failed. I excluded known probes, but internal use and unidentified tests could still be present. The activity therefore could not be treated as external customer adoption.

The inner outcome matters because a transport-level HTTP 200 can carry a tool error. To judge a successful operation, I still need to know what the user asked the agent to do. Creating a route does not establish deployment, and a successful usage read does not establish that the agent explained the response correctly.

A task review needs a known intended result and evidence concerning the relevant resource or answer. It also needs to record human intervention. A later read on the same grant can supply confirmation only if it can be associated with the resource being checked.

Accounting for concentrated activity

The inference sample at the start is an example of how one organisation can dominate a volume chart.

The sample covered selected inference routes and successful HTTP responses, with no general task-success measure. Organisations could also appear under several client categories, so summing the category totals would double-count some accounts.

An account view makes this concentration visible and allows repeat activity to be examined separately. Several activity days within a week can justify further investigation, but retention across weeks requires comparable later observations. Multiple client labels alone do not establish a qualified sales lead.

Approval events have their own limits. They record historical access decisions; current access needs an authoritative grant-state source. An approval with no later tool observation could reflect a deliberate stop, activity outside the window or a collection gap. The absence cannot diagnose abandonment.

Follow an account through a task

The account view helped me choose what to investigate. I could pick an account that returned, see what it tried and identify a particular refusal or capability to look into. I’d judge the view by whether a colleague could do the same and explain what led them there.

To find out whether an agent got useful work done, I’d then look at one task and check the answer or the resulting resource. That gives me a way to move from a concentrated volume chart to a question I can actually test. Anonymous website visits remain separate because the data I reviewed couldn’t link them to those authenticated tasks.