Your MCP tools take 15% of the context window before you type
Seven popular MCP servers expose 112 tools. Their definitions take about 29,000 tokens before the agent reads your request, roughly 15% of a 200k window. The official MCP client guide says to stop loading everything up f
Seven popular MCP servers expose 112 tools. Their definitions take about 29,000 tokens before the agent reads your request, roughly 15% of a 200k window. The official MCP client guide says to stop loading everything up front at 1% to 5%.
That's the cheap problem. The expensive one is choice: four look-alike search tools and the agent guesses from descriptions, picks the wrong source, and every call still returns 200 while the task fails.
Quick audit before adding another server:
- Run /context in Claude Code and check your tool share.
- Run /mcp and flag servers unused this week.
- Above 1% to 5%? Turn on tool search or deferred loading.
- Merge or rewrite overlapping tools.
- Pin server versions so descriptions can't change under you.
Every server should earn its keep in task success, or it's debt.
Full breakdown with the benchmark, the tool poisoning numbers and the five metrics worth tracking: https://levelup.gitconnected.com/your-mcp-setup-is-becoming-technical-debt-3917e39f9d9c?sk=73e3456a06d481d42c16c9c6b61666dd
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.