Add Blueprint Protocol to Resources — open standard for MCP server discovery

[Explorer-64]

Hi — I'd like to suggest adding Blueprint Protocol to the Resources section.

What it is

Blueprint Protocol is an open standard that solves the MCP server discovery problem. A blueprint.txt file at the root of a web app declares what the app can do, how to invoke it (MCP preferred, API fallback, UI last resort), and which tool handles each capability — in plain text, no tooling required.

Why it's relevant here

Agents connecting to tens of thousands of MCP servers burn their context window on tool definitions before reading a single user message. A blueprint lets an agent read one file and know whether a server is relevant to the task, which tool to call, and what inputs it needs — before connecting. The agent arrives with intent rather than loading everything speculatively.

The ## MCP block in a blueprint maps directly to MCP tool schemas. The ## ACCESS hierarchy treats MCP as the preferred tier over API and UI fallbacks.

Spec and examples

https://github.com/Explorer-64/blueprint-protocol

Happy to answer questions or take feedback.

[Explorer-64]

Update — v3.0.0 released:

Since posting this, the spec has moved to v3.0.0 with two changes directly relevant to the MCP context window problem:

• The blueprint now leads with a ## CAPABILITIES index so agents match their task and stop reading — no need to parse AUTH, IDENTITY, or MCP blocks unless needed
• A [MCP] flag on line one lets agents confirm MCP availability instantly
• Larger apps use Format B: an index of capability URLs with actor flags (mcp | ui | human-only). Agents fetch only the one file they need (~300-400 tokens) rather than loading the full blueprint

A weighted token cost test against v3.0.0 stubs showed 76% reduction in discovery overhead compared to no blueprint.

Live reference implementation: https://imagcon.app/blueprint.txt

Coming soon: The test harness that produced these results is being open-sourced as part of StackApps — a registry of MCP-ready indie apps. Once live, anyone will be able to run the same test against their own blueprint. Will update this issue when it's available.

[Explorer-64]

Update — benchmark results published:

Controlled discovery overhead tests are now live at https://stackapps.app/mcp-blueprint-results.

Five candidate MCP servers, fixed task, four models instructed to find the best server using the most efficient method available. No blueprint references in harness instructions — discovery was organic through signals in robots.txt and llms.txt.

Model	Without Blueprint	With Blueprint	Reduction	Notes
Claude	58 pts	13 pts	~78%	Followed signal in robots.txt
GPT-4o	58 pts	28 pts	~54%	Followed signal in llms.txt
Gemini	23 pts	9 pts	~61%	Used sitemap without blueprint; followed signal in Run B, stopped at first match
Grok	58 pts	58 pts	0%	Did not follow signal

All four models identified the correct server in both runs. Three of four showed meaningful overhead reduction through organic discovery.

Overhead is scored by weighted resource inspection cost (llms.txt: 10pts, blueprint.txt: 1pt) — weighted estimates based on typical real-world document sizes, not exact token counts.

The 0% result for Grok is an honest finding — not all models respond to the same discovery signals today. Blueprint Protocol's adoption path is the same as robots.txt and llms.txt: value accumulates as models are trained to recognise it as a standard discovery surface.

Full methodology, raw results, and test harness source: https://stackapps.app/mcp-blueprint-results