ACP – Agent Context Protocol
The intent layer above MCP — 64–97% fewer tokens
Start new thread
Shreyansh Sancheti

ACP – Agent Context Protocol - The intent layer above MCP — 64–97% fewer tokens

by•
ACP sits above MCP as an intent-resolution layer. Send intent, get back a scoped manifest: only relevant tools, auth injected server-side (never in context), execution ordering declared. Benchmark results (50 runs/scenario, tiktoken cl100k_base): • 373 → 111 tokens (-70%) for a standard query • 9,223 → 241 tokens (-97.4%) with 50 tools, 2 relevant • 1,431 → 359 tokens (-74.9%)

Add a comment

Replies

Best
Shreyansh Sancheti
Maker
📌
Built an open protocol layer that sits above MCP: instead of dumping every tool's auth tokens + schemas into the agent's context, the agent sends intent and gets back a scoped manifest — only relevant tools, auth injected server-side, execution ordering declared. Benchmark numbers (50 runs/scenario, tiktoken cl100k_base): - Standard query: 373 → 111 tokens (-70%) - 50 tools, 2 relevant: 9,223 → 241 tokens (-97.4%) - Complex multi-step: 1,431 → 359 tokens (-74.9%) Key points: - Auth never enters agent context (injected server-side) - Execution ordering is declared in the manifest, not left to the agent - It's a protocol spec (CC BY 4.0), not a library — works with any framework Go server + Python adapters for LangGraph, CrewAI, OpenAI Agents SDK. go install github.com/Clawdlinux/ninevigil-... https://github.com/Clawdlinux/ni...