How many tools can an MCP server expose before tool selection and extra context become a problem?

There is no magic maximum. The useful limit depends on the model, the quality of descriptions, how similar the tools are, and how often the agent must choose among them.

The failure mode is real, though: every exposed tool adds schema and description tokens, and similar tools make selection harder. ToolLLM used retrieval to work with a very large API collection rather than putting every API directly into one prompt. MetaTool found that models still struggle with deciding whether to use a tool and which tool to choose.

My design rule would be:

1. keep a small always-available core;
2. group tools by domain or server;
3. retrieve only the task-relevant subset;
4. make names, schemas, permissions, and failure behavior explicit;
5. log wrong-tool selections and add regression tests.

I would start with a deliberately small set and expand it while measuring selection accuracy, prompt size, latency, and wrong calls. If the agent has to choose between ten nearly identical tools, the problem is probably the interface design, not merely the number ten. Progressive discovery is safer than exposing a giant undifferentiated catalog.