Potencia tu agente
El comando propio del publicador. No es un endpoint de PlugMCP.
Host a copy with our keys and logs.
npx -y [email protected]Claude Code no tiene enlace de instalación para un comando local. Copia el comando de abajo y ejecútalo en una terminal.
claude mcp add io.github.nbhey-figma-vault -- npx -y figma-vault@0.1.5Ejecútalo en una terminal.
{ "mcpServers": { "io.github.nbhey-figma-vault": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
¿URL remota o un comando en tu máquina? En qué se diferencian las dos instalaciones
Ejecutar en PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
AbrirDel publicador
English · Русский
Figma designs pulled once into a local vault; AI agents read them offline with no rate limits
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.
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
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.
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.
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:
doc.json follows a written contract,
and a strict validator checks every file.Requires Node.js 22+.
npm i -g figma-vault
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.
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.
figma-vault add "https://figma.com/design/KEY/Project?node-id=127-4532"
# Export several selections together; links from the same file share requests:
figma-vault add "<frame link 1>" "<frame link 2>"
# Keep the node trees if image rendering is unavailable:
figma-vault add "<frame link 1>" "<frame link 2>" --no-assets
# A SECTION/CANVAS link can expand into its frames (up to 20 by default):
figma-vault add "<section link>" --expand --max-screens 20
To get the link, right-click a frame in Figma and choose Copy link to selection.
figma-vault limits "<frame link 1>" "<frame link 2>" # probe access 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> [more frame links...]
It pulls the design, reads it over MCP and builds the UI following the project's conventions.
REST API limits depend on the token, plan and seat type. Both node reads and image renders
are Tier 1 endpoints. figma-vault limits "<link>" [more links...] spends one shallow node
read plus up to ceil(S/40) screen render requests per file. It skips rendering SECTION
and CANVAS selections as whole images; their child screens are not probed. A successful
probe cannot guarantee the full pull: asset renders can require additional requests.
For a file with S selected screens and R unique raster assets, when each screen has fewer than 40 raster assets, a successful pull typically uses:
For example, two screens in one file with three unique raster assets take one node read
and one image render if the five IDs fit in a batch. At 40 or more raster assets on one
screen, screenshots are rendered separately. --expand first probes containers with
depth=1, adding at least one node request per file. SVG fallback renders, failed
requests and retries can add calls. On a real 332-node design, 141 icons needed no render calls.
After the pull, the agent never calls Figma again.
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).
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:
--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.
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.
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.
The MCP server exposes six read-only tools:
The tree is normalized:
mode: row|column with gap and padding;text.runs);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>/
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.
This is an MVP. Here is what it can't do yet.
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.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.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:
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.
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.
npm install
npm test # 85 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.
MIT
Figma designs pulled once into a local vault; AI agents read them offline with no rate limits
{ "mcpServers": { "io.github.nbhey-figma-vault": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}Herramientas
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Copia en esta página la configuración stdio para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.
Este registro no indica ninguna variable de entorno obligatoria.
Figma Vault se ejecuta localmente por stdio.
Última comprobación: 2026-11-09. Figma designs pulled once into a local vault; AI agents read them offline with no rate limits
