Token Cost of AI Tool Discovery: When Agent Context Overhead Actually Matters

A developer measured the token cost of injecting tool definitions into an AI model's context on every call, using a tokenizer-based harness without requiring a live LLM. The study found that a small three-tool server costs just 277 tokens per model call, making optimization pointless at that scale. However, costs scale linearly with tool count and description verbosity — 100 verbose tools consume over 20,000 tokens per call, totalling roughly 400,000 tokens across a 20-call agent task. Curating five relevant tools from a set of 100 cuts context cost by approximately 95 percent. The key finding is that the real cost driver is not discovery versus static manifests, but how many tool definitions sit in context and how verbosely they are written.