Skip to main contentSkip to footer
For agents

MCP setup

Connect your agent runtime to the public Blueprint MCP endpoint in three steps.

https://aidesignblueprint.com/mcp
  • Streamable HTTP
  • Protocol 2025-11-25
  • Public tools need no key

Set up in three steps

  1. Connect your client

    Add the server to your agent once. Each tab has the exact command or config; the public tools work without a key.

    Terminal

    claude mcp add aidesignblueprint --transport http https://aidesignblueprint.com/mcp

    Run in your terminal, then use /mcp inside Claude Code to sign in. No token needed: the OAuth prompt covers the paid validators.

    To use the validators, sign in: for clients other than Claude Code, create a token on Connect your coding agent.

    Not on MCP: Windsurf, GitHub Copilot, DeepSeek and Qwen read the doctrine from files; the file for each is on Start here.

  2. Check the session

    Four calls from a terminal confirm the transport, the session and a real tool before you rely on the agent. Paste them in order: the first keeps the session id in SID for the others.

    1. initialize
    2. notifications/initialized
    3. tools/call
    4. DELETE
    1. Open a session

      initialize returns the session id in the mcp-session-id header; the command keeps it in SID.

      Terminal

      SID=$(curl -s -D - -o /dev/null -X POST "https://aidesignblueprint.com/mcp" \
        -H "Accept: application/json, text/event-stream" \
        -H "Content-Type: application/json" \
        -d '{
          "jsonrpc": "2.0",
          "id": 1,
          "method": "initialize",
          "params": {
            "protocolVersion": "2025-11-25",
            "capabilities": {},
            "clientInfo": { "name": "manual-check", "version": "1.0.0" }
          }
        }' | awk 'tolower($1)=="mcp-session-id:" {print $2}' | tr -d '\r')
      echo "$SID"

      Expected

      An id of 32 hexadecimal characters. An empty line means the session did not open.

    2. Confirm the session

      The notifications/initialized message tells the server the client is ready.

      Terminal

      curl -s -o /dev/null -w "%{http_code}\n" -X POST "https://aidesignblueprint.com/mcp" \
        -H "Accept: application/json, text/event-stream" \
        -H "Content-Type: application/json" \
        -H "Mcp-Session-Id: $SID" \
        -d '{"jsonrpc": "2.0", "method": "notifications/initialized"}'

      Expected

      202
    3. Call a tool

      clusters.list is the smallest real call: if it answers, the server is reachable and the public tools respond.

      Terminal

      curl -s -X POST "https://aidesignblueprint.com/mcp" \
        -H "Accept: application/json, text/event-stream" \
        -H "Content-Type: application/json" \
        -H "Mcp-Session-Id: $SID" \
        -d '{
          "jsonrpc": "2.0",
          "id": 2,
          "method": "tools/call",
          "params": { "name": "clusters.list", "arguments": {} }
        }'

      Expected

      Four clusters: delegation, visibility, trust and orchestration.

      {"jsonrpc":"2.0","id":2,"result":{…
    4. Close the session

      DELETE with the same id ends the session on the server.

      Terminal

      curl -s -o /dev/null -w "%{http_code}\n" -X DELETE "https://aidesignblueprint.com/mcp" \
        -H "Mcp-Session-Id: $SID"

      Expected

      200
  3. Make your first call

    With the server connected, start from a real task: paste one of these prompts into your agent.

    Architecture audit

    Use the blueprint as an audit framework. List the clusters first, then propose which principles to use to assess this agent architecture and which examples to read next.

    Example lookup

    Search examples for orchestration, visibility, and steering. Group them by principle and tell me which ones are worth reading first.

    Principle explainer

    Explain the most relevant principle for this workflow with its definition, rationale, risk, and one linked example.

    Download the prompt pack

Tool reference

29 tools in four tiers. Arguments in grey, marked "?", are optional.

Prefer get and search to list. When you know the slug, principles.get(slug) and examples.get(slug) are faster and cost your model fewer tokens; when you do not, use principles.search or examples.search. Keep the list calls for browsing.

No key and no sign-in.

  • clusters.list()

    See the four clusters and the principles in each. A good first call.

  • principles.list(cluster?, lens?)

    Browse every principle, or one cluster's, with their slugs.

  • principles.get(slug, lens?)

    You know the principle's slug and want its full text.

  • principles.search(query, limit?, lens?)

    You can describe the problem but cannot name the principle.

  • clusters.get(slug)

    You know a cluster's slug and want its principles.

  • examples.search(query, principle_ids?, difficulty?, library?, limit?)

    Find code examples by topic, principle, difficulty or library.

  • examples.get(slug)

    You know an example's slug and want the whole example.

  • guides.search(query, limit?)

    Find the application guide for an engineering problem.

  • guides.list()

    Browse the application guides.

  • guides.get(slug)

    Read one application guide in full.

  • assets.list()

    Get the download links for the doctrine files and agent packs.

  • signals.feedback(task_type?, surface?, rating_clarity?, rating_usefulness?, …)

    Send feedback, only when the user asks you to.

Long runs as tasks

architect.validate can run as a task: you get an id at once and read the result when it is ready, with tasks/get and tasks/result.

Also listed on

The directories that list this MCP server (com.aidesignblueprint/blueprint) and the Python package. More appear as they go live.

Doctrine audit command

A Claude Code slash command that fetches all principles from the live MCP server, navigates key site pages with Playwright, and produces a gap report aligned to the four doctrine clusters: delegation, visibility, trust, and orchestration.
Loads all 10 principles and cluster summaries via MCP
Screenshots key pages and evaluates them against each cluster
Reports alignment, partial gaps, and missing signals with component-level references
Run with /audit-doctrine after any significant UI change
Download the command file

What happens to what I send through the MCP?

The public lookups need no key; the validators need a plan and send what you give them to OpenAI (US).

The public MCP tools are read-only lookups, except signals.feedback, which writes a record and only when you explicitly ask. signals.feedback sends its free-text fields to OpenAI (US) for triage only when permission_to_follow_up=true and a valid contact email is given, which also opens a support ticket; the contact email itself is not sent to OpenAI.
Feedback sent while signed in, or whose follow-up email matches an account, may appear, pseudonymised and shortened, in the extract our weekly lead-intelligence report sends to OpenAI (US), which is not kept in OpenAI's response store.
Every tool call writes an operational log row with what you looked up (the search text is blanked after 90 days) and a keyed one-way digest of the network address, never the address itself.
The validators (architect.validate, architect.validate_consensus, architect.certify, design.validate, spec.validate) need a plan; what you send them is processed by OpenAI (US). Retention follows your choice: standard runs persist at OpenAI, private sessions don't; either way our security and cost logs are still written, and only standard runs can be certified.
Trust and data handling