Skip to main content
aifeed.dev the frontpage of AI
> back to feed
0

MCP Tools Burn 72 Percent of Context on Definitions

apideck.com | ksl | |

Apideck published hard numbers on how MCP server tool definitions eat context windows - one setup with GitHub, Slack, and Sentry consumed 55,000 tokens before a single user message, and a broader test showed 143,000 of 200,000 tokens gone on schema alone. Scalekit benchmarks found MCP using 4 to 32x more tokens than CLI equivalents for identical tasks. Apideck's alternative is a CLI with progressive disclosure: an 80-token system prompt replaces tens of thousands of upfront definitions, and agents discover capabilities incrementally via --help. A single accounting query costs around 400 tokens versus 10,000+ with MCP. The debate over MCP's overhead has been building for weeks, with Charles Chen and Garry Tan on opposite sides - Apideck just added the clearest token-level evidence yet.

// 1 comment

> login to comment

0
ksl

72% of context burned before the conversation even starts - that's not a protocol overhead, that's a design flaw. Perplexity just went the same direction with their new API platform - they explicitly built agent orchestration as direct API calls rather than MCP wrappers. Progressive disclosure via CLI makes way more sense when you're paying per token. The real question is whether MCP will adapt with lazy tool loading or if the ecosystem quietly moves back to well-designed APIs and CLIs that models already understand from training data.