artisign
OfficialFetches font families referenced by design tokens and the Material Symbols Rounded icon font from Google Fonts, caching them locally for use in rendered screens.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@artisignCreate a login screen with our design tokens"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Artisign
An artificial artisan for UI design: agents build screens, a design system keeps them consistent.
Local-first UX design tool for AI agents. Agents design through MCP, the human reviews in the browser, state lives in plain files.
One Node process on 127.0.0.1. No database, no accounts, no hosting.
Status: Stable (1.0.0) — CLI, store, parser/model, the 24-tool MCP surface, browser preview with live reload/flows/comments, the design system (tokens, variants, drift warnings, promote_to_system), and packaging are all implemented. Artisign designs its own browser preview — see Designed with itself.
Why
Consistency across screens — the design system is first-class; screens store refs, never inline values
Token efficiency — tiered reads, field selection, diff mode on writes
Flows as first-class — click routes are data, not documentation
Render determinism — the preview renders the source, so preview equals output by construction
Related MCP server: eds-mcp-server
Install
Node.js ≥ 20.19, on macOS or Linux. Windows is not supported — the store's atomic writes (write temp file, then rename) and its path handling have never been verified there, so the tool is not shipped for it. package.json declares this via os, which makes npm install refuse the platform rather than fail somewhere later.
Two ways to run it, and the choice matters. npx artisign fetches the package on first use and needs no setup — the fastest way to try it, and enough for everything except the two screenshot tools. Installing it into a directory of your own is the other path, and the one to pick if you want screenshots:
mkdir artisign && cd artisign
npm install artisignThe commands below are written as npx artisign. With a local install, run ./node_modules/.bin/artisign instead — or npm install -g artisign and just artisign.
Canonical quickstart. The local install, from an empty directory to a connected agent. Start the daemon before registering the MCP server — the HTTP registration below runs against an already-running daemon:
mkdir artisign && cd artisign
npm install artisign
./node_modules/.bin/artisign start ./my-project
claude mcp add --transport http artisign "http://127.0.0.1:4711/mcp?project=/absolute/path/to/my-project"The first two lines repeat the install above on purpose: the point of this block is that nothing has to be assembled from elsewhere. On the npx path, drop them and start with npx artisign start ./my-project instead — the screenshot tools are then unavailable, as described below.
Optional: screenshots. get_screenshot and inspect_node need a real browser, which is intentionally not bundled (npx artisign stays light without it). Playwright is an optional peer dependency — npm neither installs it nor its ~150 MB of browser binaries unless you ask for them by name. Two ways to do that:
Next to Artisign. In the directory you installed
artisigninto:npm install --no-save playwright && npx playwright install chromium--no-savematters: it installs the package without writing it intopackage.jsonor the lockfile, which is the whole point of Playwright being optional. Without the flag npm edits both files; with some npm versions a plainnpm install playwrighteven reportsup to dateand installs nothing, because an optional peer counts as satisfied when absent.Anywhere, named by
ARTISIGN_PLAYWRIGHT_DIR. Install Playwright once into a directory of its own and point Artisign at it:mkdir -p ~/.artisign-playwright && cd ~/.artisign-playwright npm init -y >/dev/null && npm install playwright && npx playwright install chromium export ARTISIGN_PLAYWRIGHT_DIR=~/.artisign-playwright # in the shell that starts the daemonThis is the route that also works when
artisignruns from npx's cache, and the one to use when the first route leaves the repo dirty. The variable has to be set for the process that runs the daemon — for a stdio MCP server, that is theenvblock of the host's config.
artisign imports Playwright dynamically, so without the variable Node resolves it from the tree artisign itself lives in: installing Playwright into your design project has no effect. If Playwright isn't installed, or is installed but broken (a partial copy, a symlink whose target is gone), both tools fail with the fix in the error message rather than crashing the server; every other tool works without it. A Playwright installed after the daemon started is picked up on the next call; a repaired one needs a daemon restart, and the error says so.
Register the MCP server.
Claude Desktop — stdio, add to claude_desktop_config.json:
{
"mcpServers": {
"artisign": {
"command": "npx",
"args": ["artisign", "mcp", "/absolute/path/to/my-project"]
}
}
}Claude Code — stdio:
claude mcp add artisign -- npx artisign mcp /absolute/path/to/my-projectRecommended — against the running daemon (npx artisign start), streamable HTTP instead of spawning a second process. Bake the project into the URL, one stable entry per coding project:
claude mcp add --transport http artisign "http://127.0.0.1:4711/mcp?project=/absolute/path/to/my-project"The ?project= parameter scopes every tool call to that project (auto-opening it on first use), so several coding projects can share one daemon without ever restarting it. Without the parameter, calls go to the project currently active in the browser UI. init_project works even while no project is open.
Privacy. Artisign collects nothing — no telemetry, no account, no server beyond the local one it runs, no update check. Two things reach the network, neither of them about you: import_html fetches whatever URL you hand it, and the first render of a project fetches the font families named in your tokens — plus the Material Symbols Rounded icon font, always — from Google Fonts, then caches them locally (see Fonts & icons).
Quickstart
First run, in this order:
npx artisign init ./my-project # scaffold an empty project
npx artisign start ./my-project # start the daemon in the background and open the project
npx artisign status # pid, port, open projects
# then open http://127.0.0.1:4711The remaining commands, for reference — this block is a catalogue, not a sequence:
npx artisign serve ./my-project # foreground daemon (development); opens the project on start
npx artisign mcp ./my-project # stdio MCP server for Claude Desktop / Claude Code
npx artisign stop # stop the daemonartisign --help lists the flags each command accepts.
The daemon is multi-project: it runs permanently on 127.0.0.1:4711 and can hold several projects open at once. Open or create projects from the browser UI (project picker in the topbar), or let agents address any project directly via the MCP URL — no restart when you switch projects. The port comes from ~/.artisign/config.json or --port N on start and serve (a settings.port in a project's artisign.json is deprecated and ignored).
Daemon-level state — that config plus the daemon.lock holding the running pid and port — lives in ~/.artisign. Set ARTISIGN_HOME to move it somewhere else; that, together with --port, is what lets a second daemon run fully isolated from the first, which is worth doing before you try anything destructive against projects you care about.
The 24 tools
Bucket | Tool | Does |
Reads |
| The design methodology guide ( |
Reads |
| Screen list, design-system pointer, counts. Tiered, cold-start read. |
Reads |
| One screen with comment/flow indicators. Tiered + field selection. |
Reads |
| Subtree of one node, addressed as |
Reads |
| Tokens, components (with variants and the default variant's slot names), and patterns. |
Reads |
| Where-query across screens (style ref, component ref, variant, comments, text, flow). |
Reads |
| Open/resolved comments, filtered by screen or node. |
Reads |
| A mockup's variants — raw HTML, outside the ref model. |
Writes |
| Create or fully replace a screen from augmented HTML. |
Writes |
| Surgical patch by node ref or CSS selector: replace, insert, delete, set attr. |
Writes |
| Change a node's token/component/variant bindings without a full HTML parse. |
Writes |
| Design-system token mutation — one call re-resolves every bound screen. |
Writes |
| Mutate a flow edge in |
Writes |
| Screen notes/tags, design-system idea/decisions, component/pattern usage, a tag's own notes — the handoff contract. |
Writes |
| Create or revise one variant of a mockup — raw HTML, outside the design system. |
Writes |
| Read or mutate the Board's shared filter and pinned screens — daemon memory only, not available over stdio. |
Lifecycle |
| Scaffold a project directory: empty, from HTML, or from a Stitch export URL. |
Lifecycle |
| Incremental HTML ingest into an existing project, with content-hash dedup. |
Lifecycle |
| Lift an inline value or a repeated element into a token, component, or pattern. |
Lifecycle |
| Copy a chosen mockup variant into a new, design-system-bound screen (the mockup stays). |
Lifecycle |
| Delete a screen, component, pattern, or mockup; refuses a component still referenced by a screen. |
Comments |
| Answer a comment and optionally resolve it. |
Visual review |
| Screenshots a rendered screen (or one node), or one mockup variant, as a PNG — the write → screenshot → adjust loop. |
Visual review |
| Computed box and styles of one node as text — cheaper than vision for geometry questions. |
The write → screenshot → adjust loop
get_screenshot renders a screen (or, given node, crops to one element) exactly the way the browser preview does — same parse, same renderScreen — and returns a PNG, so an agent can see what a write actually produced instead of only inspecting markup. Typical use: write_html/patch_html, then get_screenshot to review, then adjust.
It needs Playwright — see the optional dependency in Install above.
scale (default 1) is the screenshot's device scale factor — 1x is enough for most visual review; go higher only when pixel-level detail matters, since a larger image costs more tokens once it reaches a model.
An Artisign project on disk
<project>/
├── artisign.json # name, version, settings (autoCommit)
├── design-system/
│ ├── tokens.json # color / spacing / typography / radius / shadow / motion
│ ├── components/<name>.html # component + variants
│ └── patterns/<name>.html # layout + interaction patterns
├── screens/<name>.html # augmented HTML, one file per screen
├── mockups/<name>/mockup.json # design explorations: variant titles/descriptions, outside the ref model
├── mockups/<name>/<variant>.html # raw HTML per variant, written verbatim
├── assets/ # local images, referenced as assets/<path>
├── flows.json # click routes between screens
├── comments.jsonl # append-only, anchored to node ids
└── .artisign/ # derived cache, gitignored — safe to delete, rebuilt on next `serve`Screens are augmented HTML — $name in class is a component ref, $name in style values is a token ref:
<button id="btn-login" class="$btn-primary" data-variant="hover"
style="padding-inline: $spacing.md" data-flow-target="dashboard">
Log in
</button>What an instance contributes is exactly what it writes: the component ref picks
the definition, data-variant picks the template, its children fill the slots,
and every other attribute on the element — style, extra classes, href,
aria-*, data-flow-target — lands on the expanded root. The instance wins
over the definition on a conflict; style and class accumulate, with the
instance's declarations last so they win the inline cascade. Only the tag is
the definition's call. The example above therefore renders the hover template
with padding-inline: 16px on top of whatever the definition sets.
Everything is human-readable and diffable. With autoCommit on (the init default), every write op becomes one git commit named after the tool and its target (write_html: dashboard, set_tokens: color.primary, ...) — that is the audit log.
Browser preview
Served at http://127.0.0.1:<port> once the daemon is running — plain ES modules, no build step:
Project picker — switch between open projects in the topbar; open an existing folder or create a new project through an in-app folder browser (no restart needed). With nothing open, an empty state offers both actions plus recent projects.
Screens — sidebar screen list, rendered source (preview = output, by construction)
Board — every screen laid out as a tile on one scrollable surface, with flow edges drawn between them (toggleable), a 5%–200% zoom slider (Fit all, 100%, or Ctrl/Cmd+wheel/pinch to zoom around the cursor), and the sidebar's shared filter plus any screen pinned onto the board regardless of the filter, so a flow is something you see rather than something you reconstruct
Elements panel — the selected screen's nodes listed by the id an agent addresses them with
Notes panel — the screen's own notes plus any tag notes that apply to it, both rendered as Markdown
Flow mode — click an element with
data-flow-targetto jump to the screen it points atComment mode — click an element to anchor a comment; appended to
comments.jsonl, picked up bylist_comments/ answered viareply_commentDesign System — token swatches by bucket, every component with every variant rendered side by side
Mockup view — a mockup's variants rendered live, side by side in columns, each with its title and description
Follow mode — off by default; the topbar's Follow toggle (three states: off / Following / Follow paused) makes the preview track a live agent session's own MCP tool calls, switching the view/selection to whatever it touches and briefly outlining the affected nodes (solid teal for a read, dashed violet + wash for a write, or a dashed border + "Updated" badge around the whole screen after
write_html). Every tool call also lands in the right panel's new Activity tab (last 50, in-memory, newest first) regardless of follow's own state. Human navigation — a screen/tab click, clicking an entry in the Activity feed, or opening comment/inspect mode — pauses it; only the toggle's own "Resume" un-pauses itLive reload over SSE — any file change on disk shows up in the browser within ~200 ms
Deterministic render baseline
Every rendered surface — screens pane, Design System variants, and (later) get_screenshot — wraps the rendered fragment in the same small baseline document, so what an agent designs is what renders everywhere, pixel for pixel. The guarantee: box-sizing: border-box on every element, zero body margin, antialiased text. A screen root declared width: 390px with padding measures exactly 390px — no implicit UA margin or box-model surprise to account for.
Fonts & icons
Any font family named in design-system/tokens.json's typography bucket (e.g. "600 24px/1.2 'Plus Jakarta Sans', sans-serif") is fetched once from Google Fonts as woff2 and cached under .artisign/fonts/ — offline-first after the first serve. A fetch failure (no network, unknown family) is silent: the render falls back to the system font, it never blocks.
Icons are Material Symbols Rounded, always cached alongside your typography fonts, used via ligature names:
<span class="icon">close</span>Never substitute a Unicode glyph for an icon — the .icon class and the ligature name are what make the rendered result deterministic.
Images
Drop a local image into a project's assets/ folder (subfolders are fine) and reference it as a project-relative path, assets/<path>, in an src attribute or a CSS url():
<img src="assets/hero.png">
<div style="background-image: url(assets/icons/logo.svg)"></div>The reference stays portable in the source file. At render time it resolves two ways, the same split webfonts use: a page loaded over HTTP (the preview iframe, /api/render/*) gets it rewritten to /api/assets/<path>?project=<path> unconditionally, tagged with the project it was rendered from so a subresource request out of an iframe's srcdoc still resolves against the right project even with several open at once — a path that doesn't resolve to a file under assets/ still gets the URL, and 404s there, rather than being checked against the filesystem twice; a headless render with no HTTP base URL (get_screenshot, inspect_node) gets it inlined as a data: URI instead, and there an unresolvable path is left as-is. Either way, the browser's own broken-image handling makes the problem visible rather than the image silently vanishing.
Repo structure
src/cli/ CLI entry point — init, start, stop, status, serve, mcp
src/daemon/ daemon lifecycle — project registry, global config/lock (~/.artisign)
src/init/ project scaffolding (init) and artisign.json
src/store/ Store interface, FsStore, watcher, atomic writes, git auto-commit
src/model/ HTML parser, canonical ref model, index
src/tools/ the 24 tools (reads, writes, lifecycle, comments, visual review)
src/mcp/ MCP server — stdio + streamable HTTP
src/http/ internal JSON API + SSE, consumed by the preview
src/preview/ browser preview — plain ES modules, no build step
design/ the tool's own UI, designed with the tool (see below)
scripts/ one-off dev tooling, not shipped in the npm packageDesigned with itself
Artisign's own browser preview is designed in Artisign. design/ is a real
project in the on-disk format described above — tokens, components, screens, flows — and
it is the source the interface you see at 127.0.0.1:4711 was designed from.
Every token that has a counterpart in the shipped src/preview/style.css :root block —
all 19 colours, the six spacing steps, radius.pill and shadow.menu — mirrors it by value
and by role name, so those cannot quietly drift apart: change one and the mismatch is
visible in the other. The typography scale and radius.sm/radius.md have no custom
property behind them, so they are design decisions rather than a mirror. Values with no
counterpart at all stay literal rather than implying a binding that does not exist. Open it
the same way as any other project:
npx artisign serve designThat makes it the largest worked example in the repo, and the honest answer to whether the tool is good enough to build with — it built this one.
Development
npm install
npm run type-check # tsc --noEmit
npm run lint # eslint
npm test # vitest run
npm run build # compile src/ to dist/
npm run dev -- serve ./my-project # run the CLI from source via tsxOnly dist/, src/preview/ (minus its test files), assets/ and the two guide documents in docs/ ship in the published package — see files in package.json. prepack runs the build automatically before npm pack/npm publish, so the tarball can't ship a stale dist/.
Documentation
What | Where |
Design methodology for agents (normative, served by | |
The same workflow long-form, for humans and skill authors | |
The tool's own UI, designed with the tool |
|
Issues & security
Bug reports, feature ideas and questions go to the
issue tracker — pick the matching
template. Vulnerabilities go through private reporting instead, never a public
issue: see SECURITY.md. Code contributions are currently not
accepted; CONTRIBUTING.md explains why and what is welcome.
License
Copyright (C) 2026 Christian Körbs
Artisign is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, version 3.
Artisign is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License along
with this program. If not, see https://www.gnu.org/licenses/. The full text is
in LICENSE.
Two things the license does not cover:
Your designs are yours. The AGPL covers Artisign's own source code, not the screens, components, tokens or projects you create with it. Files Artisign writes into your project folder belong entirely to you.
The name is not licensed. "Artisign", its logo and its domains are not part of the AGPL grant. Forks may use the code but need their own name.
This server cannot be deployed
Maintenance
Related MCP Connectors
UI design from prompts, screenshots, and URLs for AI coding agents and theme tokens.
Give your agent a real design system: tokens, measured WCAG contrast, and rules to follow.
Agent-Native design tool - create and edit visual designs with agent assistance
Brand-safe MCP for AI agents to create editable, on-brand graphics and automate variants.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceMCP server that exposes your design system components and tokens to AI agents, preventing duplicate component creation and hardcoded token values.5 npm9MIT
- AlicenseBqualityBmaintenanceEnables AI agents to access design system tokens and component contracts through MCP, reducing token usage and ensuring consistency.29MIT
- AlicenseBqualityCmaintenanceEnables AI agents to directly control Figma Desktop via MCP, supporting UI creation, editing, prototyping, and variable management with over 60 tools.65873 npm1MIT
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to remotely operate visual design tools via MCP protocol, with composable architecture for image processing, file operations, and workflow automation.MIT