VQV Signal

AI AT WORK WATCH TECHNICAL

Deferred MCP loading reportedly cut a Claude Code setup's tool overhead from 141k to 1.3k tokens

According to a post on X by @RajaPatnaik, three MCP servers were adding about 141,000 tokens of tool schemas to every Claude Code request. Deferred loading reduced that to about 1,300 tokens, and he said coding tasks on GLM 5.3 were about 62.5% cheaper on average. He says these are his own measurements from his setup. The available excerpt is cut off, so it does not include further detail on his method.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-20T02:40:00+00:00 Detected 2026-11-20T10:01:17+00:00
View original source

According to a post on X by @RajaPatnaik, three MCP servers were adding about 141,000 tokens of tool schemas to every Claude Code request. Deferred loading reduced that to about 1,300 tokens, and he said coding tasks on GLM 5.3 were about 62.5% cheaper on average. He says these are his own measurements from his setup. The available excerpt is cut off, so it does not include further detail on his method.

AI-assisted summary based on the listed source.

Raja Patnaik said three MCP servers were adding about 141,000 tokens of tool schemas to every Claude Code request, and deferred loading cut that to about 1,300, making coding tasks on GLM 5.3 about 62.5% cheaper on average. He warned that Claude Code and Codex turn deferred loading off when pointed at a...

Developers who run Claude Code or Codex with several MCP tool servers may see large differences in per-request token use and cost, depending on how tools are loaded. He warned that both tools turn deferred loading off when pointed at a third-party open-model host, so users on those hosts should check their setups. Watch whether similar savings appear in other configurations, since the figures come from one user's own measurements.

Teams using AI at work may want to compare this against current productivity and review workflows.

Signal Strength 81% Technical label WATCH Public Interest 51 Category AI AT WORK Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 82 Practical Impact Score 28 Novelty Interest Score 70 Consequence Score 22 Curiosity Score 16 Shareability Score 64

VQV surfaced this signal because it is recent, relevant to AI Coding Tools, connected to Scobleizer News Pulse (X).