i'm an indie app builder and vibe coder. i've shipped over 30 small business apps — invoicing, inventory, packing slips, tax tracking. and now apparently an open standard for ai agents.
that last one surprised me too.
the problem i kept running into: even the best ai agents hallucinate when they look at a web app. they guess where to click. they call the wrong tools. they fail silently and confidently. i had 30 apps that existed — agents just couldn't see them.
so i built blueprint protocol. a blueprint.txt file at the root of your app that tells agents exactly what it can do, which tools to call, what's human-only, and how to complete each flow step by step. no mcp required. it works for any agent trying to navigate any web app.
then i ran a benchmark and found out it also cuts mcp discovery overhead by 78%.
what it does for mcp specifically
mcp has two calls. tools/list returns every tool the server exposes. tools/call calls a specific tool by name.
without a blueprint, an agent calls tools/list first, reads through all N tool definitions, picks one, then calls it.
with a blueprint, the agent already knows the tool name and exact parameters. it calls tools/call directly. tools/list never happens.
that's not just cheaper discovery. it's the removal of an entire round trip.
the benchmark
i ran a controlled test across four models — claude, gpt-4o, gemini, and grok — five candidate mcp servers, one correct match. agents were told to find the best server efficiently. no blueprint references in the harness instructions. discovery was organic.
model without blueprint with blueprint reduction
claude 58 pts 13 pts ~78%
gpt-4o 58 pts 28 pts ~54%
gemini 23 pts 9 pts ~61%
grok 58 pts 58 pts 0%
overhead scored by weighted resource inspection cost — llms.txt at 10pts, blueprint.txt at 1pt, reflecting typical real-world document sizes.
grok's 0% is an honest result. not all models follow discovery signals today. that's expected — robots.txt and llms.txt didn't work everywhere on day one either.
v3.0.0 — agent-first structure
the spec just hit v3.0.0. two changes that matter for mcp:
[MCP] flag on line one — agents confirm server availability instantly
format b capabilities index — root file is ~150 tokens, agents fetch only the one capability file they need (~350 tokens). total: ~500 tokens vs thousands from tools/list
try it
spec:
benchmark results: https://stackapps.app/mcp-blueprint-results
open standard. mit licensed. no tooling required. works with or without mcp. if you have a web app or an mcp server, a blueprint takes about 10 minutes.
SOCIAL SHARE CARD GENERATOR