Skip to main content
Past about thirty tools, servers increasingly stop shipping the whole list. A lightweight classifier reads the opening message and registers only the tools plausibly relevant to that intent. As one author put it: you have quietly turned a routing problem into a retrieval problem. That is often the right trade. It also breaks two things MCPulse reports, and the new failure it introduces is silent by construction — the right tool was never a candidate, nothing was called, and nothing distinguishes that from nobody having asked.

​
1. Schema cost stops being one number

Schema size assumes the whole tool list ships on every connection. Under dynamic registration it does not: the payload is whatever the classifier selected, and it differs per session. MCPulse reports actual per-session schema bytes rather than a single startup figure, and flags when the two differ materially — so you know which number you are looking at. A server registering 12 of 120 tools is paying a tenth of the bill the startup figure would suggest, and being told the larger number would be wrong in your favour, which is the worse direction.

​
2. Dead tools becomes ambiguous

“Never called” now means one of two things with completely different fixes: MCPulse distinguishes them by recording which tools were present in each session’s list, not just which were called. Without that, a rewrite gets recommended for a tool the model never saw — advice that cannot possibly work.
This needs no configuration. The startup payload already carries the tool list per session; under dynamic registration it simply carries a different list each time, and that difference is the signal.

​
3. Retrieval failure rate

The number that says whether dynamic registration is net positive or just relocating the problem: how often the classifier excludes a tool that would have been the right answer. You cannot currently see this at all. The failure is silent — no wrong call is made, because no call is made. Measuring it needs ground truth: a labelled set of intent-to-tool pairs, replayed against the classifier. That is the same artefact the contrastive-prompt harness produces, and it is the reason that harness is worth building before you need it.

​
Collisions get worse here, not better

The intuition runs the other way, so it is worth stating plainly. If your classifier reads tool descriptions to decide what to load, then two tools a model cannot tell apart are two tools a classifier cannot tell apart. The ambiguity does not go away — it moves one layer earlier, to a place where it produces no wrong call to point at. A collision at selection time shows up as a wrong tool in your call log. The same collision at retrieval time shows up as nothing at all. So the schema checker matters more under dynamic registration, not less. It reads the descriptions your classifier is reading.