figma-vault

figma-vault

English · Русский

Pull a Figma frame once. Then Claude Code, Codex or any MCP agent reads it as often as it likes: offline, with no Figma token and no rate limits.

• 3 Figma requests per screen, spent once when you pull it. After that it's zero, however many times the agent looks.
• Icons cost nothing. A real screen had 141 icons, and they took no extra requests.
• Commit the vault and the team needs no Figma access. No licence, no token.

Try it in 30 seconds. You don't need a Figma token for this:

npm i -g figma-vault
figma-vault init && figma-vault demo && figma-vault check

How it works: the pull normalizes the frame and writes it to disk. The agent then reads it over MCP.

one developer:      figma-vault add <link>  →  vault  →  git commit
the rest of the team:  the agent reads it over MCP, no Figma token needed

The problem

Figma's MCP server and REST API are rate-limited, and on some seats the limit is tiny. While building a UI, an agent goes back to the design all the time to check a margin, a colour or a font size. That can mean dozens of calls in one session. You hit the limit fast and the work stops.

With figma-vault, the number of Figma calls doesn't depend on what the agent does. Each design costs a fixed number of calls, spent once when you pull it.

	Figma Dev Mode MCP	figma-vault
who calls Figma	the agent, while it works	the CLI, once
how many calls	grows with the agent's work	fixed per design
does every reader need a token	yes	no
works offline	no	yes

There's a side effect: the design becomes a file in your repository. A developer without access to the Figma file (no licence, no permissions, no token) still gets the whole design.

How it compares

Other projects work on the same problem. The Framelink Exporter plugin exports a design to a zip for offline use, and there are caching forks of Framelink's Figma MCP. figma-vault takes a different approach:

• The vault is meant to be committed. One person pulls a design and the whole team reads it from git. Nobody else needs Figma access.
• Icons cost zero requests. Vector paths come with the node tree, and the SVGs are built locally. Rendering them through Figma would eat the image-render limit.
• The format is fixed and validated. doc.json follows a written contract, and a strict validator checks every file.

Install

Requires Node.js 22+.

npm i -g figma-vault

Option 1. Per project

figma-vault init

This creates .figma-vault/, registers the server in the project's .mcp.json (other entries are left alone), adds .env to .gitignore and installs a /figma slash command for Claude Code.

Use --no-command to skip the /figma command, so the agent gets only the read-only vault_* tools.

Option 2. Nothing in the repository

figma-vault init --global

The vault lives in ~/.figma-vault/ and the server is registered in the agent's user config. The project repository never learns the tool exists. Use this when you can't add config or dependencies to the project.

Usage

figma-vault add "https://figma.com/design/KEY/Project?node-id=127-4532"

To get the link, right-click a frame in Figma and choose Copy link to selection.

figma-vault limits "<link>"   # check your Figma rate limits BEFORE pulling
figma-vault list              # what has been pulled so far
figma-vault check             # verify the whole chain works
figma-vault demo              # add a demo design, no token needed

CLI output, errors and the MCP tool descriptions are in English by default; the /figma command it installs is in English too. README.ru.md is a translation of the documentation only.

After init, Claude Code has one command:

/figma <frame link>

It pulls the design, reads it over MCP and builds the UI following the project's conventions.

What a pull costs in Figma requests

REST API limits depend on the token, the plan and the seat type, and you can't look them up in advance. Run figma-vault limits "<link>" before your first pull. It makes the same two requests a pull would and tells you whether the pull will go through.

A full pull of one screen costs three requests. We measured this on a real design with 332 nodes:

	requests
node tree (/v1/files/:key/nodes)	1
frame screenshot (/v1/images)	1
raster images (/v1/images, in batches of 40)	1 per 40 images
icons and vectors	0

On that design, the 141 icons cost zero requests. So the cost of a pull doesn't depend on how many icons the design has. After the pull, the agent never calls Figma again.

Check the result against the design

figma-vault verify compares the page the agent built with the design. The package ships no browser: the snapshot is taken with whatever the project already uses (Playwright, DevTools, the agent's browser tool).

• Mark the root element of the page with data-figma-node-id="<root node id>", and the main blocks with the ids of their nodes.
• figma-vault verify <docId> --snippet prints a script. Run it on the page and save what it returns to a file.
• figma-vault verify <docId> --snapshot <file> checks it:
  • texts: every visible text of the design must be on the page;
  • geometry: each marked block must match the design's position and size, within 2 px by default (--tolerance).

The result is PASS (exit code 0), FAIL (1, with a list of what is off) or INCOMPLETE (2, when nothing is marked and geometry can't be checked). The /figma command runs this step itself when the project has a browser.

Try it without a Figma token

You only need a token to pull new designs. Reading works without one, which is the normal setup for everyone except the person who pulls. So you can test the whole install without creating a token:

figma-vault init && figma-vault demo && figma-vault check

check starts the real MCP server as a child process and talks to it over the protocol. It exits with code 1 when something is actually broken, so you can use it in CI.

Token

In Figma, go to Settings → Security → Personal access tokens → Generate new token. The token needs only one scope: File content → Read-only.

Put it in a .env file in the directory you run the CLI from, as a line FIGMA_TOKEN=figd_..., or set it as an environment variable. If both are set, the environment variable wins.

What the agent gets

The MCP server exposes six read-only tools:

tool	returns
vault_list	the designs in the vault
vault_get_doc	a design's node tree; maxDepth limits the depth
vault_get_node	the subtree of one node
vault_search	nodes found by name or text
vault_get_tokens	colours, typography, effects
vault_get_asset	an image or an icon

The tree is normalized:

• auto-layout becomes mode: row|column with gap and padding;
• coordinates are relative to the parent;
• styles, gradients and text come in one predictable shape, including differently coloured fragments inside one text (text.runs);
• nodes hidden in Figma are kept with hidden: true. By default the agent doesn't get them, but a parent reports how many it has (hiddenOmitted). Pass includeHidden: true when building a reusable component whose optional slots are switched off on this screen.

The raw Figma response is saved next to it in raw.json for debugging.

On a real design, normalization shrinks the tree about 7 times:

raw.json   1.9 MB   raw Figma response
doc.json   256 KB   what the agent reads

This saves more than disk space. An agent that reads 256 KB instead of 2 MB rebuilds the UI noticeably more accurately.

Vault layout

<vault>/
  index.json                 list of designs
  <docId>/
    doc.json                 normalized tree, the main artifact
    raw.json                 raw Figma response
    screenshot.png           render of the whole frame
    assets/                  icons and images

The figma-vault/doc@1 schema is described in docs/CONTRACT.md (in Russian). A strict validator checks it. Vaults pulled with 0.1.0 (doc@0) are still read as they are; there is no need to pull them again.

doc.json stands on its own: the MCP server never reads raw.json and never touches the network.

Known limitations

This is an MVP. Here is what it can't do yet.

• Image-render limit. Figma's /v1/images endpoint runs out sooner than tree reads do, and the block can last for days. Icons no longer depend on it, but the frame screenshot and raster images still do. The client retries up to three times, following Retry-After, and never sleeps for more than a minute. If Figma asks it to wait for days, the pull keeps the node tree, records the reason as a warning and doesn't fail. Run figma-vault limits before pulling.
• Only colour and weight are kept for fragments inside a text. A gradient, a different font or an underline on part of a text is lost.
• Designs with many component variants make a large doc.json. Hidden slots are kept now. On a 332-node screen that added 3,400 hidden nodes, and the file grew from 444 KB to 5 MB. The agent still gets about 100 KB by default, but the file goes into your git.

How it was tested

The fixture in the repository is synthetic: 29 nodes, enough to install the tool and confirm the chain works without a Figma token.

We checked accuracy separately, on two real designs whose content isn't published here. Results on a product screen with 332 nodes, 15 levels deep:

check	result
design texts present in the rebuilt markup	41 of 41
icons	15 of 15, 0 broken links
placeholders needed	1 (a raster avatar)
Figma requests to pull	3
doc.json against the contract	passes

Not checked: pixel-perfect match. We compare with Figma's reference render by eye, using npm run dev → /compare/<name>. figma-vault verify checks texts and the geometry of marked blocks, but not colours, fonts or pixels.

What it deliberately doesn't do

It isn't a SaaS, a Figma replacement or a real-time sync. There is no auth, billing, teams or users, and no server side. The MCP server runs on the developer's machine as a child process of the agent and talks over stdio. There is nothing to deploy.

More docs (in Russian)

• docs/INTEGRATION.md: rolling it out to a team, the no-trace mode, checking without a Figma token.
• docs/ADAPT.md: adapting the tool to a project's design system.

Development

npm install
npm test          # 63 tests
npm run typecheck
npm run dev       # local preview of rebuilt markup and comparison with the design

Two AI agents write this project in parallel, following the protocol in AGENTS.md: separate areas of the code, a task board and a shared log. Each commit is tagged with its author.

License

MIT