geny-svgforge 0.1.0

AI가 좌표 대신 의미(spec)만 내면 결정론적 레이아웃으로 깨끗한 다이어그램 SVG를 만드는 라이브러리

geny-svgforge

Give an AI a semantic spec instead of raw coordinates, and a deterministic layout engine produces a clean diagram SVG with zero overlaps or clipping.

When an LLM writes <svg> by hand, it can't reliably reason about text widths, box bounds, curve paths, or the viewBox — so elements overlap and captions get clipped. geny-svgforge inserts a layout layer between the AI and the SVG. The AI emits only a JSON spec ("two rows of labeled tokens, posN labels, connect A's token 2 to B's token 3 with a curve, a side note, a caption") — no coordinates. The library measures text with real font metrics, sizes every box, routes connectors around obstacles, and fits the viewBox to the content. Overlap and overflow become structurally impossible.

Same lineage as edit2ppt (AI emits PPT structure, engine renders) and Contextifier (structure-preserving document parsing).

---

Install

pip install geny-svgforge            # core (SVG)
pip install 'geny-svgforge[png]'     # + PNG export (cairosvg)
pip install 'geny-svgforge[mcp]'     # + MCP server

Quickstart (Python)

from geny_svgforge import render

spec = {
    "type": "token-sequence",
    "title": "Absolute position changes, relative pattern remains",
    "rows": [
        {"label": "sentence A", "tokens": [
            {"text": "나는", "pos": "pos 0"},
            {"text": "오늘", "pos": "pos 1"},
            {"text": "밥을", "pos": "pos 2", "id": "a2"},
            {"text": "먹었다", "pos": "pos 3", "variant": "highlight"},
        ]},
        {"label": "sentence B", "tokens": [
            {"text": "나는", "pos": "pos 0"},
            {"text": "정말", "pos": "pos 1", "variant": "accent"},
            {"text": "오늘", "pos": "pos 2"},
            {"text": "밥을", "pos": "pos 3", "id": "b3"},
            {"text": "먹었다", "pos": "pos 4", "variant": "highlight"},
        ]},
    ],
    "connectors": [{"from": "a2", "to": "b3", "color": "accent"}],
    "note": {"title": "What the model must learn",
             "lines": ["Memorizing absolute positions is brittle to length changes.",
                       "Relative distance and surrounding patterns are handled in attention."]},
    "caption": "Same relative token relationship survives an absolute shift",
}

result = render(spec)        # portable SVG with the used glyphs embedded
print(result.warnings)       # []  ← no overlap / no clipping (lint passed)
open("out.svg", "w").write(result.svg)

CLI

geny-svgforge render spec.json -o out.svg
geny-svgforge render spec.json -o out.png      # PNG (requires [png])
geny-svgforge validate spec.json               # validate before rendering
geny-svgforge schema -o schema.json            # dump the JSON Schema

MCP server

geny-svgforge ships an MCP server over stdio so any MCP-compatible agent can request diagrams. After pip install 'geny-svgforge[mcp]' the server is launched with:

geny-svgforge-mcp                 # console script
# or
python -m geny_svgforge.mcp_server

Tools exposed

Tool	Input	Returns
get_diagram_schema	–	JSON Schema describing the spec (the agent learns the format from this)
validate_diagram_spec	spec	{ ok, errors[], warnings[] } — check before rendering
render_diagram	spec	{ svg, width, height, warnings[] } — if warnings is non-empty, fix the spec and call again

Client configuration

Add the server to your MCP client config. The standard shape is an mcpServers map keyed by a server name.

Claude Desktop (claude_desktop_config.json), Cursor (~/.cursor/mcp.json), or Claude Code (.mcp.json):

{
  "mcpServers": {
    "geny-svgforge": {
      "command": "geny-svgforge-mcp"
    }
  }
}

Zero-install with uv (no prior pip install needed):

{
  "mcpServers": {
    "geny-svgforge": {
      "command": "uvx",
      "args": ["--from", "geny-svgforge[mcp]", "geny-svgforge-mcp"]
    }
  }
}

Claude Code can also add it from the CLI:

claude mcp add geny-svgforge -- uvx --from 'geny-svgforge[mcp]' geny-svgforge-mcp

A typical agent flow: call get_diagram_schema once to learn the format → emit a spec → call render_diagram → if warnings is non-empty, repair the spec and retry.

---

How it works

Three layers: Spec (JSON Schema) → Layout Engine → Renderer.

• Real font metrics — text width is computed in pixels by summing glyph advances via fontTools (no browser, no headless engine). The same font used for measurement is embedded into the SVG, so measured layout == rendered output.
• Deterministic layout — boxes are sized to their text, connectors are routed through the inter-row band away from boxes, and the viewBox/padding is derived from the bounding box of every element — so nothing can clip.
• Font embedding — only the glyphs actually used are subset and inlined as a base64 @font-face, so the SVG renders identically everywhere (browsers, resvg). to_png() renders via the installed font (raster_safe) because cairosvg ignores embedded @font-face.
• Lint — a post-layout pass flags box overlaps and canvas overflow. It should always be empty; if not, the warnings are returned to the agent so it can fix the spec.

Diagram types

Type	Description
token-sequence	Rows of position-labeled token boxes, with inter-row connectors, a side note, and a caption

(flow, grid, stack, callout, … are planned — the spec is extensible via the type field.)

Roadmap

• Automatic collision resolution (constraint / force based)
• Visual self-repair loop: render → rasterize → multimodal critique → fix spec → re-render
• More diagram types, themes, templates, accessibility (<title>/<desc>/aria)

License

MIT. Text is measured against — and a subset is embedded from — a system font (e.g. SIL OFL Noto Sans CJK). When embedding, the embedded font's own license applies.