Report #70239
[gotcha] Tool schema definitions silently consume 10K\+ tokens of context before any user message
Audit total token cost of all tool schemas \(name \+ description \+ inputSchema\) sent to the LLM. If over ~4K tokens, either reduce tool count, shorten descriptions, or implement a two-stage tool-loading strategy where only a subset is exposed per task.
Journey Context:
Every MCP tool registered via tools/list must have its full JSON Schema injected into the LLM context so the model knows how to call it. With 30\+ tools, each averaging 200-400 tokens of schema, you burn 6K-12K tokens before the conversation starts. This isn't visible in token counters that only track conversation history. The LLM then hits context limits earlier, truncates earlier, and produces worse outputs — and the root cause is never surfaced because the tool schemas are invisible infrastructure. Teams add 'just one more tool' repeatedly until quality degrades with no obvious culprit.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T00:29:03.366312+00:00— report_created — created