sap-arch-mcp

MCP server that exposes the sap-architecture skill pipeline as tools, so an LLM agent can generate SAP Architecture Center-style .drawio diagrams using the full deterministic pipeline (template scaffolding, icon extraction, autofix, validation, corpus scoring) instead of fetching raw files per request.

Tools exposed

Tool	Purpose
`list_templates`	Browse/filter the 71 bundled official SAP templates
`scaffold`	Rank templates against a request, copy best match, return XML + design recipe (mandatory first step)
`extract_icon`	Official BTP service icon → ready-to-paste mxCell (32×32, grid-snapped)
`extract_asset`	Any starter-kit asset: connectors, area shapes, numbers, brand names
`autofix`	Mechanical fixes: grid, hex case, absoluteArcSize, strokeWidth, fonts
`validate`	Full SAP-style validator (bent arrows, palette, overlap, pill verbs...)
`score`	Corpus similarity / SAP-likeness gate (PASS ≥ 90)
`compare`	Fingerprint diff against one named template
`render_png`	Headless PNG export (optional, needs drawio + xvfb)
`get_reference_doc`	Read skill reference docs (gotchas, palette, layout...)

Quick start (Docker, recommended)

# docker-compose.yml
services:
  sap-arch-mcp:
    image: python:3.12-slim
    container_name: sap-arch-mcp
    restart: unless-stopped
    ports:
      - "8600:8600"
    volumes:
      - ./sap-architecture:/opt/sap-architecture:ro
      - sap-arch-work:/work
    environment:
      SKILL_DIR: /opt/sap-architecture
      WORK_DIR: /work
    command: >
      sh -c "pip install --quiet 'mcp[cli]' mcpo &&
             mcpo --host 0.0.0.0 --port 8600 --
             python3 /opt/sap-architecture/mcp-server/server.py"
volumes:
  sap-arch-work:

git clone https://git.alexzaw.dev/alexz/sap-architecture.git
docker compose up -d

Put the container on the same Docker network as OpenWebUI if you want to reference it by service name.

Wire into OpenWebUI

Admin Settings → Tools → Add Connection → http://sap-arch-mcp:8600 (or http://<host>:8600). The tools appear automatically; mcpo serves OpenAPI docs at /docs.

Optional: PNG rendering

Inside the container (or bake into an image):

apt-get update && apt-get install -y wget xvfb libgbm1 libasound2 libnss3 libgtk-3-0
wget -O /tmp/drawio.deb https://github.com/jgraph/drawio-desktop/releases/download/v28.0.6/drawio-amd64-28.0.6.deb
apt-get install -y /tmp/drawio.deb
printf '#!/bin/bash\nexec xvfb-run -a /usr/bin/drawio "$@" --no-sandbox --disable-gpu\n' > /usr/local/bin/drawio-headless
chmod +x /usr/local/bin/drawio-headless
export DRAWIO_CLI=/usr/local/bin/drawio-headless

Without it, everything except render_png works — the validator and scorer are the real quality gates.

Agent workflow (what the LLM should do)

scaffold(request) → pristine SAP template XML + design recipe
Surgical relabel of the XML (never draw from scratch)
extract_icon / extract_asset for any new cards
autofix(xml) → validate(xml) → fix errors → repeat
score(xml) → PASS ≥ 90, else compare against the source template and restore drift
Deliver .drawio (+ render_png if requested), narrate the flow below the diagram

3.2 KiB Raw Blame History Unescape Escape