The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →No: the Model Context Protocol (MCP) specification does not set a universal 90-tool ceiling. It supports paginated tool discovery, so a server can return its catalog across multiple pages. A limit or failure can instead come from a particular client’s pagination behavior, the host’s handling of the catalog, pressure on the model’s context, or the model’s ability to choose the right tool. Those are separate problems, and “90+ tools” is not enough to identify which one is occurring.
What does MCP itself limit?
MCP tool discovery uses tools/list. The official Tools specification, version 2026-07-28, says a server that declares the tools capability must respond with the tools currently available to the requesting client. It supports pagination and caching; it does not prescribe a maximum number of tools such as 90.
As an Amazon Associate I earn from qualifying purchases.
The available set may change over time and may vary according to authorization. The specification says it must not vary per connection or as a side effect of other requests on that connection. This describes protocol behavior, not a promise that every host will display, load, or send every listed tool to a model in the same way.
Where can a large catalog run into trouble?
“The catalog breaks” can describe several different failures. Identifying the layer matters: a server response problem is not the same as a client stopping pagination or a model selecting the wrong tool.
#1 Best Overall
| Layer | What may fail | What the evidence establishes |
|---|---|---|
| MCP server and protocol | A list request or page traversal may fail or return incomplete results. | The specification supports paginated discovery but does not establish a 90-tool ceiling or a particular server’s behavior. |
| Client or SDK | The client may stop aggregating pages, impose a safeguard, or handle caching in a particular way. | The TypeScript SDK documents a default 64-page cap for one automatic aggregation method; that is not a tool-count limit. |
| Host and model context | A host may register or expose only some tools, or a large set of descriptions may become difficult to use within the model’s context. | Filtering and retrieval are documented architectural approaches; no universal context threshold is established by the cited material. |
| Model tool selection | The model may choose an unsuitable tool even when discovery and registration worked. | A reported study finds selection accuracy below 90% at particular tool counts for particular models; it does not establish a universal breaking point. |
What the TypeScript SDK’s 64-page setting means
The MCP TypeScript SDK documentation says listTools() without a cursor walks pages and returns an aggregate. Its automatic aggregation path has a default listMaxPages limit of 64. Calls made with an explicit cursor fetch individual pages without that automatic-aggregation cap.
That number counts pages, not tools. The cited documentation does not establish how many tools a page contains, so it cannot be converted into a maximum tool count. If a client’s result stops early, check whether it uses this aggregation method, whether it follows returned cursors, and whether it applies a separate host-level limit. Do not infer a 90-tool protocol limit from a client’s page safeguard.
Rank #2
Does model tool selection get worse with more tools?
It can, but the available numerical evidence is specific rather than universal. A 2026 arXiv paper, MCP Server Architecture Patterns for LLM-Integrated Applications, reports selection accuracy below 90% at 10–15 tools for Claude Haiku 4.5 and at 20–30 tools for Sonnet 4. These are reported results for the study’s settings; they are not MCP limits, and they do not show that every model or tool set fails at those counts.
Free tools Windows power users keep installed
One-click scans. No signup required.
Tool count alone is an incomplete measure. Descriptions, overlap between tool purposes, the task mix, and the model and host configuration all matter when evaluating whether selection is reliable. The cited result should be treated as a reason to test a real workflow, not as a universal cutoff.
How can an implementation keep a large catalog usable?
One option is to reduce the tools registered or presented at a given time. AWS guidance describes both application-side filtering and semantic search as ways to narrow the tools an agent uses. These are architecture patterns, not guarantees that any specific host will implement them or that they will improve every task.
- Filter by context: expose only tools relevant to the user, task, permissions, or active workflow.
- Retrieve on demand: search the broader catalog and make a smaller set of relevant tools available for the current request.
- Keep discovery behavior explicit: verify that the client follows pagination and understand any page-aggregation cap or cache behavior.
- Evaluate the actual task mix: measure whether the model selects the right tool, not merely whether the server can list the catalog.
- Test catalog changes: check that tools added, removed, or made available through authorization are reflected as expected by the client and host.
These approaches address different constraints. Filtering can reduce what the model must consider, while pagination addresses how a client retrieves a long list. Neither substitutes for checking that the server returns the expected pages or that the model selects correctly.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to diagnose a reported “90+ tool” failure
- Locate the first missing or incorrect result. Determine whether the server fails to return a page, the client stops following cursors, the host registers fewer tools than expected, or the model sees the tools but chooses incorrectly.
- Inspect pagination separately from counts. Record the pages requested and returned, whether a next cursor is present, and whether the client uses automatic aggregation or explicit cursor calls. A page limit is not a tool limit.
- Check host and authorization behavior. Compare the tools available to the requesting client with the tools the host registers. The protocol permits availability to depend on authorization.
- Test selection with representative tasks. If all tools are present but the wrong one is selected, evaluate the model against the actual descriptions and task mix. A successful
tools/listresponse alone does not demonstrate good tool choice. - Apply the fix at the failing layer. Correct pagination if retrieval is incomplete; adjust host aggregation if it truncates; filter or retrieve tools on demand if the active catalog is unwieldy; improve or evaluate selection if the model chooses poorly.
What can be concluded about 90 tools?
There is no official MCP specification figure establishing a 90-tool cap. The protocol supports paginated discovery; a documented TypeScript SDK safeguard is expressed in pages, and the reported model-selection findings are tied to particular models and study settings. Without naming a specific host, version, and observed failure, “the catalog breaks at 90” is not a general MCP rule.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
