How many tokens does your MCP setup burn before you've typed anything? A DIY measurement
LitheShade1X7 · reddit · 2026-10-22
After threads about context bloat, the author finally measured their own MCP setup: start a fresh chat with all servers on, send a one-word message, check the token count, then repeat with servers off one at a time. The total was far larger than expected, and a few servers with huge tool lists accounted for most of it.
Mitigations now in use:
- Enable servers per project instead of keeping everything on
- Drop servers whose tools overlap
- Prefer servers where the agent searches for a tool first instead of loading every schema
The search approach seems cleanest for large catalogs, at the cost of an extra round trip before the first real call. The post includes a reusable measurement method.
More from coding & agent
- Rescale built an AI triage agent in 5 days with Strands Agents and Claude on Bedrock — AWS ML Blog · 2026-10-23
- I Let AI Commit to My Repo—and I'm Still Deciding Where My Line Is — petrusenko_max · 2026-10-23
- Free visual layout editor for MOSS-Image ships on Hugging Face Spaces — Subject_Sir_2796 · 2026-10-23
- Dev discovers "random 256-bit number" was implemented as random 0-256, yielding tiny outputs — prateekj · 2026-10-23
- Airbnb's CTO details Project Everest, a "company brain" giving AI agents institutional knowledge — samcharrington · 2026-10-23
- DiffResearch: Diffusion LM Backbone Cuts Deep Research Latency While Improving Reports — ITMO · 2026-10-23