Ink/Stitch MCP Bridge
Enables AI assistants to convert SVG designs into machine-embroidery files (DST, PES, etc.) without Inkscape, and provides tools for lettering, validation, and previews.
README
Ink/Stitch MCP Bridge
An MCP server that gives AI assistants full headless access to Ink/Stitch machine-embroidery tooling — no Inkscape installation required.
SVG designs go in; stitch-ready machine files (DST, PES, JEF, EXP, VP3, and ~17 more), PNG previews, validation reports, and digitized lettering come out.

The image above is the bridge's own output: generate_lettering("Hello", "Amitaclo") →
machine-ready DST → render_png_preview.
MCP client (Claude Code / Claude Desktop)
│ stdio (JSON-RPC)
▼
inkstitch-mcp server (FastMCP, this package)
├── native tools ──────────► pystitch + lxml (stats, PNG render, format
│ conversion, param editing, font/format lists)
└── subprocess tools ──────► .venv-inkstitch python
vendor/inkstitch/inkstitch.py --extension=…
(export, import, lettering, validation,
stitch-plan previews, transforms)
Tools (21)
| Tool | What it does |
|---|---|
health_check |
Verify toolchain; optional end-to-end smoke export |
list_formats |
Machine formats Ink/Stitch reads (~50) and writes (~22) |
list_fonts |
140 bundled digitized lettering fonts |
list_known_params |
All legal per-element stitch parameters (~136) |
list_elements |
Inventory an SVG: ids, labels, stitch type (fill/stroke/satin), params |
get_element_params / set_element_params |
Read/write inkstitch:* stitch settings on elements |
export_embroidery |
SVG → machine file (format from output extension) + stitch stats |
import_embroidery |
Machine file → editable stitch-plan SVG |
export_zip |
One SVG → many formats in a ZIP, optional panelization |
design_stats |
Stitch/jump/trim/color counts, threads, size (mm) of a machine file |
convert_machine_format |
Machine file → machine file (e.g. DST → PES) |
render_png_preview |
PNG image of the stitches (from machine file or SVG) |
thread_list |
Plain-text thread/color report |
troubleshoot |
Structured validation: errors/warnings with positions & fixes |
density_map |
Stitch-density heat-map layer (thread-break risk) |
preview_stitch_plan |
Render the actual stitch plan into an SVG layer |
cleanup_document |
Remove problem-causing tiny objects (dry-run by default) |
transform_elements |
stroke_to_satin, auto_satin, auto_run, fill_to_stroke, jump_to_stroke, break_apart, outline |
remove_embroidery_settings |
Strip embroidery params/commands from an SVG |
generate_lettering |
Text → embroidered lettering in any bundled font |
Setup
Requires Python 3.11 (Ink/Stitch's pinned dependencies do not support newer interpreters) and git. Developed and tested on Windows 11; the code paths are OS-aware but only Windows has been exercised.
# --recursive pulls the Ink/Stitch submodule AND its fonts submodule (140 fonts)
git clone --recursive https://github.com/GreerBK/InkStitchMCP
cd InkStitchMCP
py -3.11 -m venv .venv-inkstitch
# inkex's repo has a test file exceeding Windows path limits; scope the fix to this install:
set GIT_CONFIG_COUNT=1&& set GIT_CONFIG_KEY_0=core.longpaths&& set GIT_CONFIG_VALUE_0=true
.venv-inkstitch\Scripts\python -m pip install -r requirements-inkstitch.txt
.venv-inkstitch\Scripts\python -m pip install -e .
If you cloned without --recursive: git submodule update --init --recursive.
Verify everything with the health_check tool (or run the test suites below).
Paths are overridable via INKSTITCH_MCP_ROOT / INKSTITCH_MCP_REPO /
INKSTITCH_MCP_PYTHON / INKSTITCH_MCP_OUTPUT (see src/inkstitch_mcp/config.py),
plus INKSTITCH_MCP_TIMEOUT / INKSTITCH_MCP_SLOW_TIMEOUT (seconds) and
INKSTITCH_MCP_TRACE=1 for stderr tracing of subprocess calls.
Registering with a client
Claude Code — either copy .mcp.json.example to .mcp.json (auto-loads for this
project) and fill in your path, or register globally:
claude mcp add inkstitch -- <path-to-InkStitchMCP>\.venv-inkstitch\Scripts\python.exe -m inkstitch_mcp
Claude Desktop — add to %APPDATA%\Claude\claude_desktop_config.json:
{
"mcpServers": {
"inkstitch": {
"command": "<path-to-InkStitchMCP>\\.venv-inkstitch\\Scripts\\python.exe",
"args": ["-m", "inkstitch_mcp"]
}
}
}
Tests
.venv-inkstitch\Scripts\python tests\e2e.py # 34 checks, every tool, real designs
.venv-inkstitch\Scripts\python tests\protocol_smoke.py # real MCP stdio round-trip
Design notes (hard-won)
These are the traps this bridge encodes; they matter if you modify runner.py:
- Exit codes lie. Ink/Stitch exits 0 for parse errors, "no embroiderable elements", and domain errors — with the explanation on stderr and an empty payload. Success is judged per-tool from payload presence, never from the return code.
- stdout is binary. Embroidery/ZIP payloads come over stdout (the Inkscape
convention); the
outputextension switches stdout toO_BINARYon Windows. Always capture bytes; stderr is UTF-8, flushed once at process exit. stdin=DEVNULLis load-bearing. Under an MCP client, the server's stdin is the live JSON-RPC pipe; letting a child inherit it deadlocks the child interpreter at startup on Windows (diagnosed via py-spy: zero CPU, no Python frames).- Legacy SVGs summon an invisible dialog. Every extension's
load()runs the version migrator with a blocking wx prompt for unversioned files containing Ink/Stitch attributes — including--extension=update_svgitself. The preflight inrunner.pydetects the condition (same XPath aslib/update.py) and migrates via_migrate_helper.py, which calls the migrator directly with the prompt disabled. - Never expose raster preview modes.
png_realistic,png_simple, realistic-DPI stitch plans, and the ZIP--format-png-realistic/simpleflags all shell out to aninkscapebinary. PNG previews here use pystitch's native renderer instead. - The
outputextension's parser is hand-rolled. Options must be single--name=valuetokens,--formatis mandatory, values containing=crash it, and--outputis swallowed (hence stdout capture there,--output=<tmpfile>elsewhere). - Arg spelling is per-extension. Some use hyphens, some underscores, some both
(
batch_letteringmixes them in one parser). Flags are hardcoded per tool from the vendored argparsers — don't "normalize" them. --del_paramsmust always be passed toremove_embroidery_settings: its argparse default is the Python boolTruewithtype=str, which crashes when omitted.batch_letteringdumps the input SVG on validation failure with exit 0 — the ZIP magic (PK) check is what actually detects failure.- Documents parameterized by this bridge are version-stamped.
set_element_paramswrites the currentinkstitch_svg_versionmetadata; without it, the next tool call would run the v0→v4 legacy migration on freshly written modern params and silently change their meaning (e.g. solid strokes forced tozigzag_stitch). - Slow tools run off the event loop. Subprocess-backed tools are registered via
blocking_tool(async +anyio.to_thread); registering them as plain sync tools freezes all MCP traffic for the duration of an export. custom_file_nameis a path fragment inside the zip extension — it is validated against a strict character allowlist to prevent writing outside the target directory.
License
This bridge code: MIT (see LICENSE). Ink/Stitch is referenced as a git submodule —
not distributed in this repository — and is GPL-3.0-or-later; pystitch and other
dependencies carry their own licenses. If you distribute a bundle that includes the
vendor/inkstitch checkout, GPL terms apply to that component.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。