
Localfig
io.github.LucasGorgalPublished by GitHub account LucasGorgalv0.4.1Updated Oct 24, 2026
The Figma desktop app as an MCP server: full Plugin API for AI agents. Local, no token, no quota.
Overview
Localfig exposes the Figma desktop app to an AI assistant through a local plugin bridge, giving full Plugin API read and write access to the open file.
- What you can do
- Check plugin connection, current file, page, selection and frames, and run arbitrary Plugin API JavaScript in the open file via helpers for text, auto-layout, tokens and components. Inspect structure: metadata dumps with styles, bound variables and computed CSS, plus variable collections, paint, text and effect styles. Search layers and text across pages, and review what changed since the last call or save named snapshots and undo the last tool call. Work with team libraries: list variables, import components, styles and variables, and place instances. Export nodes to PNG, JPG, SVG, PDF or JSON, and place local images into fills.
- When to use it
- Useful when an assistant should read, edit or generate real Figma designs in the desktop app without a cloud token, for example building on an existing design system, extracting tokens and metadata for code handoff, or exporting renders and assets.
- Requirements
- Node.js 18 or newer, the Figma desktop app (dev plugins cannot be imported in the browser), and an MCP client that can launch a stdio server. Setup writes a plugin to a local folder that must be imported into Figma once, and the plugin must be run and left open in the target file each session. Optional environment variables set the home folder, bridge port and export folder.
Installation
In SourceWeft
- Open Localfig in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
localfig
The Figma desktop app as an MCP server. localfig gives an AI agent the full Figma Plugin API on the file you have open — through a tiny dev plugin and a loopback HTTP bridge, entirely on your machine.
- Reads and writes. Frames, text, auto-layout, variables, components, image fills, exports, undo — anything the Plugin API can do, the agent can do.
- No Figma cloud API, no personal access token, no per-account quota.
- Zero dependencies. Node built-ins only. Nothing leaves your machine.
- Works with any MCP client that can launch a stdio server: Claude Code, Cursor, Windsurf, Cline, Claude Desktop and more.
Why localfig
Requirements
- Node.js 18 or newer
- The Figma desktop app — dev plugins cannot be imported in the browser
- An MCP client: Claude Code, Cursor, Windsurf, Cline, Claude Desktop, …
Install (once)
That generates your local secret, writes the plugin to ~/.localfig/plugin/, registers localfig with every MCP client it finds on your machine, and prints the manifest path for the next step. --dry-run shows what it would do without touching anything, and --clients=cursor,vscode limits it to the clients you name. Prefer a clone? git clone https://github.com/LucasGorgal/localfig.git && cd localfig && node server.mjs --setup does the same.
A client counts as installed when its config folder exists. Setup merges into existing files instead of replacing them: other servers and settings stay, an existing localfig entry keeps its env, and a .localfig.bak copy is written before any change. Files it cannot safely rewrite, such as a settings.json with comments, are left untouched and reported with the entry to paste. Restart any client that was already open.
For a client that is not in the list, add this to its MCP configuration:
On Windows, use "command": "cmd", "args": ["/c", "npx", "-y", "localfig"] instead: most clients start servers without a shell, and npx is a .cmd script there.
Then import the plugin into Figma (once):
- Open any design file in the desktop app. The Plugins menu only exists inside a file, not on the home screen.
- Figma menu (the logo, top-left) → Plugins → Development → Import plugin from manifest… — or press
Ctrl/(Cmd/) and type import plugin from manifest. - Pick
~/.localfig/plugin/manifest.json(the setup output prints the full path).
The manifest asks for two permissions — teamlibrary (library components and tokens) and currentuser (who is driving) — and registers a relaunch button.
Use (every session)
- Open the file you want to work on.
- Run the plugin:
Ctrl/→ localfig — orCtrlAltP(CmdOptP) to re-run the last plugin, or the Reconnect localfig button in the properties panel of a file it has run in. A small panel appears and says Connected to localfig once the MCP server is up. Leave it open. - Ask your agent to do Figma work. Every tool acts on the file where the plugin is running — there is no file key, which is why nothing touches the cloud.
Closing the panel (or the file) disconnects; re-run the plugin to reconnect. This is the one manual step: Figma provides no way to launch a plugin from outside.
Tools
figma_eval environment
Code runs inside the plugin sandbox, wrapped in an async function: top-level await and return both work.
Globals: figma and helpers.
helpers.createText({characters, font: {family, style}, fontSize, color: '#hex', width, lineHeight, letterSpacing, textCase, textAlign, name, parent, x, y})— loads the font and ordersresize()beforetextAutoResize.helpers.setText(node, chars)— loads the node's real fonts, then sets characters.helpers.loadNodeFonts(node).helpers.set(node, props)— batch assign;layoutModefirst,width/heightviaresize, sizing modes after.helpers.rgb('#rrggbb'),helpers.rgba('#rrggbbaa')— hex to the{r, g, b[, a]}fills want.helpers.reveal(nodes)— scroll and zoom the viewport to what you just made.helpers.notify(msg)— a toast.helpers.collection(name, [modes]),helpers.token(collection, name, type, value | {mode: value}),helpers.bind(node, 'fills' | 'strokes' | prop, variable)— create design tokens and bind properties to them, one line each.helpers.importComponent(key),helpers.instance(key, parent, props)— team-library components.helpers.query(root, sel)— minimal selectors:TEXT,[name=X],[name*=X],A B,a, b.helpers.createAutoLayout(dir, props).helpers.command(kind, payload)— run any typed command (tokens,metadata,find,changes,export…) from inside a script:(await helpers.command('tokens', { collection: 'Brand' })).result.
Plugin API gotchas the helpers protect you from
resize()on a TEXT node resetstextAutoResizetoNONE; on an auto-layout frame it sets both sizing modes toFIXED. Resize first, then set the sizing.- Return plain data from
figma_eval. Real Figma nodes are collapsed to{id, name, type}so a stray node reference cannot flood the result — usefigma_metadatato inspect a subtree. - Load fonts before any text mutation.
figma.notify()works here (the cloud MCP blocks it).
Design-system awareness
figma_tokens returns the file's variables the way a design system thinks about them:
figma_metadata with styles: true then reports, per node, boundVariables (which token drives each property), fillStyle / textStyle / effectStyle names, and for instances the main component and its componentProperties; css: true adds the CSS Figma computes for each node. Together they let an agent build on a design system instead of hard-coding values — and hand code off from a real file.
Images
- In:
figma_place_imagesends the file's bytes to the plugin UI, which is a full Chromium iframe. It decodes any browser-supported format, downscales tomaxSide(default 4096, Figma's hard limit) with high-quality smoothing, converts non-native formats to PNG, and hands the result to the main thread. The tool result reports what happened (prepared). - Out:
figma_exportwrites files to~/.localfig/exports/and attaches PNG/JPG (≤ 2 MB each) to the result as images, so the agent sees the render without another call.JSONexports the subtree in Figma REST API shape.
Undo, versions and change tracking
Every mutating tool call is checkpointed with figma.commitUndo, so it lands in the file's undo history as one step a person can Ctrl+Z (the most recent call is committed 60 s after it finishes, or at the next call). Within that window figma_history {action: "undo"} reverts the last call from the agent side — triggerUndo reverts to the last checkpoint — and {action: "snapshot"} saves a named version before risky edits. The plugin also records node changes on every page it has seen (Figma delivers them batched, after the fact): figma_changes returns what changed since a seq, with edits caused by tool calls tagged byPlugin by a time-window heuristic and hidden by default — the agent can notice what the person changed in Figma between calls instead of overwriting it.
Multiple sessions
The first localfig process owns the bridge port; later ones detect it via /health, run as clients, and proxy through it — same tools either way. If the owner dies, the next tool call on a client re-binds the port and takes over ("re-election" in the log); the plugin re-registers when it reconnects.
Files and configuration
node server.mjs --help lists the flags.
Security
The bridge binds loopback only (127.0.0.1 and ::1). Every endpoint except /health requires a 32-hex secret generated on first run into ~/.localfig/token and baked into ~/.localfig/plugin/ui.html. Edit plugin/ui.template.html in the package, never the generated ui.html. Nothing is sent to any server.
Design notes
- Why a plugin? It is the only local surface with full read/write access to an open Figma document. The REST API is cloud, needs a token, and cannot write nodes.
- Why HTTP long-poll and not WebSocket? The plugin main thread has no network at all; only the UI iframe does. Long-poll needs no dependency and no framing code, and
http://localhostis exempt from mixed-content blocking. Figma's manifest validator rejects raw IPs innetworkAccess, hencelocalhost— and the IPv6 mirror listener, because Windows resolveslocalhostto::1first. - Binary stays binary on the plugin channel. Exports leave as
POST /blob; images enter viaGET /assetinto aUint8Array. Only JSON goes through the command channel. (Renders are base64-encoded once, in the MCP result, where the protocol requires it.) - Why a home directory? The package may live in the npx cache or a read-only global install; the plugin folder you import must not move when the package updates, and per-user state must not live in a package.
Platform notes
Developed and tested on Windows 11 with the Figma desktop app. macOS and Linux should work unchanged — Node built-ins only, no native code — but have not been exercised yet. Reports and fixes welcome.
Test
Spawns the server in an isolated home, speaks MCP over stdio, and impersonates the plugin over HTTP — covering the handshake, all ten tools, file round-trips, inline images, owner re-election, and setup against sandboxed MCP client configs, without Figma running. 55 assertions.
Layout
License
MIT — see LICENSE.
Source: README.md at commit 0242238
Tools
0Version history
1- v0.4.1LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts