archdiagram

Turn an architecture description into an editable diagram whose nodes are the official vendor service icons (Azure App Service, AKS, Key Vault, S3, GKE, ...) - not a flowchart of labelled boxes.

A coding agent (Cursor, Claude Code, Codex) translates a natural-language architecture into a small JSON spec; the deterministic Python core renders that spec to:

Lucid note: Lucid does not map draw.io's named vendor stencils or its img/lib/... icon references, so those render blank on import. .drawio/.vsdx therefore embed icons (--drawio-mode portable) to maximise render success, and the PDF is the reliable icons-always-render deliverable for Lucid.

Gallery

Rendered from the specs in examples/ - same engine, real vendor icons.

Related MCP server: dgmo-mcp

Design constraints

• Core engine = pure Python stdlib. No graphviz, no browser, no third-party Python deps. Spec validation is hand-rolled (no jsonschema). .vsdx is written with zipfile + xml.

• Python 3.11+. Tests use stdlib unittest.

• Optional, isolated, non-stdlib deps:

  • mcp SDK - only inside mcp_server/.

  • Node + @resvg/resvg-js - rasterise icons to PNG; used by .vsdx (embedded glyphs) and .pdf. If absent, .vsdx falls back to labelled boxes and .pdf errors with an install hint.

  • Node + pdfkit + svg-to-pdfkit - SVG -> PDF for .pdf.

• Vendor icons are not committed. Fetch them with tools/download_icons.py.

Quick start

# 1. (optional) fetch vendor icons into the local cache
python -m tools.download_icons --vendors azure,aws,gcp,kubernetes

# 2. (optional) install the Node rasterise/pdf bridge
cd node && npm install && cd ..

# 3. render a spec
python -m archdiagram.cli render examples/sample_spec.json --format pdf    -o out/arch.pdf
python -m archdiagram.cli render examples/sample_spec.json --format drawio -o out/arch.drawio --drawio-mode portable
python -m archdiagram.cli render examples/sample_spec.json --format vsdx   -o out/arch.vsdx

Spec format

{
  "title": "Web platform",
  "direction": "LR",
  "groups": [{ "id": "app", "label": "Application tier", "vendor": "azure" }],
  "nodes": [
    { "id": "web", "service": "azure.app_service", "label": "Web App", "group": "app" },
    { "id": "aks", "service": "azure.aks",         "label": "AKS",     "group": "app" },
    { "id": "kv",  "service": "azure.key_vault",   "label": "Key Vault" }
  ],
  "edges": [
    { "source": "web", "target": "aks", "label": "http" },
    { "source": "aks", "target": "kv" }
  ]
}

service is "<vendor>.<service_key>". Unknown services degrade gracefully to a labelled box.

Layout

Node positions are computed by a deterministic, hand-rolled layered layout (no graphviz). Provide explicit x/y on a node to override the computed position.

Use from an AI agent

This repo is built to be driven by agentic AI tools (Cursor, Claude Code/Desktop, Codex, ...) straight from Git. Pick either path:

1. As a skill

Clone the repo and point your agent at skill/SKILL.md. It teaches the agent the spec schema, catalog lookup, and render commands so it can go from a natural-language architecture to rendered files. See also AGENTS.md.

git clone https://github.com/bharathgnana/archdiagram.git
cd archdiagram && python -m tools.download_icons

2. As an MCP server

pip install "archdiagram[mcp]"
python -m mcp_server.server          # stdio transport

Exposes validate_spec, render_diagram, search_catalog, and list_catalog tools. A sample client config is in examples/mcp.json (drop it into .cursor/mcp.json or claude_desktop_config.json).

Repository layout

archdiagram/        pure-Python core (spec, layout, catalog, emitters)
mcp_server/         optional MCP server (only place the `mcp` SDK is used)
node/               optional Node bridge (resvg rasterise + svg->pdf)
tools/              icon downloader + icon source map
skill/              SKILL.md for agentic AI tools
examples/           sample specs + mcp.json
tests/              stdlib unittest suite

Tests

python -m unittest discover -s tests

Contributing

Anyone can clone, install (pip install -e .), and use it. To propose changes, fork and open a pull request - see CONTRIBUTING.md for the dev setup and workflow.