yasuda-figma-mcp

English | 日本語

A self-hosted, read-only Figma MCP server for GitHub Copilot (agent mode) in GitHub Codespaces. It renders screenshots locally inside your running Figma app — exactly like right-click → Copy as PNG — and never uploads anything to a public S3 URL.

Why this exists. The official Figma MCP / REST image endpoints return rendered images on a public, unauthenticated S3 URL (figma-alpha-api.s3..., valid up to 30 days). Anyone with the URL can view your design. For security-sensitive orgs that's a non-starter. This server avoids it entirely: image bytes are produced by the Figma client's local exportAsync and streamed straight to Copilot — no S3, no Figma REST, no API token, no outbound HTTP at all.

Verified on real hardware (2026-06-17): real Codespace + real Figma desktop. Copilot discovered all 9 tools; get_screenshot rendered a 6576×3952 frame and a FigJam board locally (no S3); every tool returned real data over a private gh codespace ports forward tunnel with no public port.

How it works

 your laptop                                    your Codespace
┌────────────────────┐   private gh tunnel     ┌──────────────────────────┐
│ Figma desktop app  │   (GitHub-authed,       │ bridge  (:3055)          │
│  └ plugin ─ ws://localhost:3055 ════════════► │   ▲                      │
│                    │   no public port)       │ MCP ─stdio─► Copilot      │
└────────────────────┘                         └──────────────────────────┘
   exportAsync() renders locally → bytes never leave this path

• bridge (src/bridge.ts) — a token-authenticated WebSocket relay that pairs the plugin with the MCP server. Forwards messages; persists nothing.

• mcp (src/mcp.ts) — the stdio MCP server Copilot launches. Exposes the read tools; talks only to the bridge over localhost.

• plugin (plugin/) — runs in your Figma app, executes each read op against the currently open file, and returns data/bytes.

The plugin reaches the Codespace bridge through gh codespace ports forward — a private, GitHub-authenticated tunnel to your own localhost. No port is ever made public.

📐 Deep dive: the full request lifecycle, wire protocol, target resolution, and exactly how each tool works (Figma APIs, inputs, outputs) are documented in docs/ARCHITECTURE.md.

Related MCP server: figma-opencode-mcp

Tools (9, all read-only, all local)

If you don't pass url/nodeId, the tool operates on your current selection in Figma.

🤝 Teach Copilot to use these tools well: a ready-made /yfigma prompt for GitHub Copilot lives in integrations/ (design→code workflow, token reuse, node targeting).

Setup ⭐ recommended (npx + Codespaces secret)

No clone, no build, no per-project config. Do the one-time setup once; after that each project costs only tunnel + Connect.

One-time setup (per developer)

1 — Generate a token (any random secret string; this gates the bridge channel):

openssl rand -hex 24
# → copy the output, e.g. 7f3a9c4e…  (used in steps 2 and the plugin)

2 — Register it as a Codespaces user secret (recommended — one value for all your Codespaces, nothing to manage per project):

• GitHub → Settings → Codespaces → Secrets → New secret

  • Name: BRIDGE_TOKEN

  • Value: the token from step 1

  • Repository access: the repos you'll use it in (or All repositories)

• …or via CLI: gh secret set BRIDGE_TOKEN --user --app codespaces (paste the value when prompted)

Every Codespace you open now has $BRIDGE_TOKEN injected automatically; the MCP and embedded bridge read it with no prompt.

3 — Add the MCP server to your VS Code user config (once; applies to every workspace and Codespace):

• Command Palette → "MCP: Open User Configuration"

• paste this and save:

{
  "servers": {
    "yasuda-figma-mcp": {
      "command": "npx",
      "args": ["-y", "github:Rikuto-des/yasuda-figma-mcp", "mcp"],
      "env": {
        "BRIDGE_EMBED": "1",
        "BRIDGE_PORT": "3055",
        "BRIDGE_URL": "ws://127.0.0.1:3055",
        "BRIDGE_CHANNEL": "default"
      }
    }
  }
}

BRIDGE_EMBED=1 makes the MCP host the bridge in-process, so there's no separate bridge to start. (Local VS Code, no Codespaces secret? Add "BRIDGE_TOKEN": "<your-token>" to this env block instead.)

Codespaces note: user-level config only reaches a Codespace if VS Code Settings Sync is on. The reliable option is to commit the same { "servers": { … } } as .vscode/mcp.json in each repo you'll use it in — then it's always present when that repo's Codespace opens. (Not .github/, not devcontainer.json — those don't define MCP servers.)

4 — Give the local gh CLI the codespace scope (needed for the tunnel; once):

gh auth refresh -h github.com -s codespace

5 — Get the Figma plugin (once) — see Getting the plugin. Recommended: publish it org-internal so it shows up in everyone's plugin list with no manifest import.

Per session (each project, each time)

Order matters — the bridge must be listening on :3055 inside the Codespace before the tunnel can connect: (1) bridge up in the Codespace → (2) tunnel from your own computer → (3) plugin Connect.

1 — Open the project's Codespace and make sure the bridge is listening on :3055. With the MCP configured (above) and BRIDGE_EMBED=1, Copilot starts the MCP — which hosts the bridge — when the MCP first loads (open Copilot Chat in agent mode, or run "MCP: List Servers" → Start). A fresh Codespace builds the package once (~30–60 s).

Simplest way to be sure (and to test the path): in the Codespace terminal, start it explicitly and leave it running:

echo "$BRIDGE_TOKEN"                                   # must be non-empty
npx -y github:Rikuto-des/yasuda-figma-mcp bridge       # → [bridge] listening on 0.0.0.0:3055

2 — Open the tunnel on YOUR OWN computer's terminal — NOT the Codespace terminal. It maps your computer's localhost:3055 to the Codespace so the Figma desktop app can reach it:

gh codespace list                                # find your Codespace name
gh codespace ports forward 3055:3055 -c <name>   # leave this running
# (from a local clone of this repo you can also run:  npm run tunnel)

connect failed (Connection refused) here = nothing is listening on :3055 in the Codespace yet → do step 1. No Codespace found = you ran this inside the Codespace → run it on your own computer instead.

3 — Run the Figma plugin ("Yasuda Figma MCP") on your computer. It auto-connects with your saved settings (first time: paste the token — reveal it with echo "$BRIDGE_TOKEN" in the Codespace terminal). Status turns Ready.

4 — Use Copilot (agent mode), e.g. "screenshot my current Figma selection". The 9 yfigma_* tools are available.

That's the whole per-project cost: bridge up → tunnel → plugin Connect.

Alternative: run from a clone of this repo

1. Open a Codespace on this repo (Code → Codespaces → Create). The devcontainer runs npm install && npm run build && npm run setup, which auto-generates your personal token and writes .env.

2. Start the bridge in the Codespace terminal:

   npm run bridge        # prints: [bridge] listening on 0.0.0.0:3055

3. Open the tunnel on your local machine (needs the gh CLI with the codespace scope — run gh auth refresh -s codespace once):

   gh codespace ports forward 3055:3055 -c <your-codespace-name>
   # (with a local clone you can just run:  npm run tunnel )

4. Run the plugin in the Figma desktop app (see install options below) → paste your token → Connect (the token is printed by npm run setup; re-run it or grep BRIDGE_TOKEN .env to see it again).

5. Use Copilot (agent mode). It auto-starts the MCP server and discovers the 9 tools — no token prompt. Try: "screenshot my current Figma selection".

2nd time onward (resume)

Codespaces auto-stop when idle. To resume: reopen the Codespace → npm run bridge → tunnel → run the plugin → Connect. Your token persists in .env.

Getting the plugin

Recommended — publish it once as an org-internal plugin (Figma Organization/Enterprise):

1. In Figma desktop: Plugins → Development → Import plugin from manifest… → select plugin/manifest.json (one admin does this).

2. Plugins → Development → Manage plugins in development → Yasuda Figma MCP → Publish.

3. Set visibility to "Only available to your organization", add a name/description, and upload an icon (use plugin/icon.svg, exported to a 128×128 PNG). Publish.

4. Members now run it from Plugins → (your org's plugins) — no manifest import needed. To ship changes, re-publish.

Fallback — manifest import (any Figma plan): each developer imports plugin/manifest.json via Plugins → Development → Import plugin from manifest… (requires Figma desktop; the browser app can't import dev plugins).

Token (per user)

The token is a random secret that gates who can join the bridge channel (defense in depth on top of the private tunnel). Neither method needs a Copilot prompt — the MCP reads BRIDGE_TOKEN from the environment.

• Codespaces user secret — recommended ✅ — one value, set once, inherited by every Codespace across all repos; nothing to manage per project. GitHub → Settings → Codespaces → Secrets → BRIDGE_TOKEN (or gh secret set BRIDGE_TOKEN --user --app codespaces). Reveal it in a Codespace with echo "$BRIDGE_TOKEN".

• .env — per-Codespace alternative — used by the clone-this-repo flow. npm run setup generates a token and writes .env (gitignored); npm run bridge and the MCP load it via --env-file-if-exists. Rotate: rm .env && npm run setup.

Paste the same value into the Figma plugin once. If both a secret and a matching .env exist, the secret (the process environment) wins.

Security model

• No public port. The bridge is reachable only through the private, GitHub-authenticated gh tunnel to your own localhost. Nothing is exposed to the internet.

• No S3, no Figma REST, no token, no outbound HTTP. Images come from the local exportAsync; all other data from the Figma plugin API. (grep the repo: there is no fetch, no api.figma.com, no /v1/images.)

• Defense in depth. Private tunnel (who can reach the port) + per-user BRIDGE_TOKEN (who can join the channel) + plugin networkAccess restricted to ws://localhost:3055. The bridge persists nothing.

• Stop the tunnel (or the Codespace) and the path disappears.

Limitations (honest)

• Only the file you currently have open in Figma — the plugin can't reach files you aren't viewing (the official MCP/REST can fetch any file by key).

• get_design_context returns raw design data, not Figma's opinionated code generation — your model writes the code.

• get_libraries sees only team-library variable collections (plugin-API limit), not full component-library enumeration.

• Code Connect is out of scope (not reachable from the plugin API).

Scripts

Troubleshooting

• No Codespace found (running tunnel) — you ran the tunnel inside the Codespace. It must run on your own computer's terminal (it bridges your localhost to the Codespace). Check there with gh codespace list.

• connect failed (Connection refused) on the tunnel — the tunnel is fine, but nothing is listening on :3055 in the Codespace. Start the bridge there: open Copilot so it launches the MCP (embedded bridge), or run npx -y github:Rikuto-des/yasuda-figma-mcp bridge in the Codespace terminal. Then retry the tunnel.

• echo "$BRIDGE_TOKEN" is empty in the Codespace — the secret isn't present. A Codespaces secret added after the Codespace was created only appears after a rebuild / reopen (or export BRIDGE_TOKEN=… temporarily).

• "Figma plugin is not connected" — run the plugin and Connect; check the token/URL/channel match. The bridge log should show plugin joined.

• "Bridge is not connected" — is the bridge (or Copilot's MCP) running in the Codespace? Is BRIDGE_TOKEN set?

• Plugin won't connect — is gh codespace ports forward 3055:3055 running on your own computer? Plugin URL ws://localhost:3055? Tunnel dropped → restart it.

• Connection drops intermittently — the plugin↔bridge hop crosses the gh tunnel, so two things help: (a) the plugin sends a keepalive every 20s and the bridge tolerates a missed pong (built in — make sure you have the latest plugin + package), and (b) run a persistent standalone bridge instead of the embedded one. With BRIDGE_EMBED=1, Copilot stopping/restarting the MCP when idle also drops the bridge → the plugin drops. Fix: remove BRIDGE_EMBED from your mcp.json env, and run npx -y github:Rikuto-des/yasuda-figma-mcp bridge in a kept-open Codespace terminal. The MCP then reconnects to it over localhost (instant) without touching the plugin's tunnel connection. Also confirm the tunnel is still running and the Codespace hasn't idle-stopped.

• Multiple pages vs files — switching pages in the same file needs nothing: the plugin stays connected and follows the active page (target other pages by nodeId/url, or allPages:true for search). Switching files closes the plugin (Figma runs plugins per file) — just re-run it; it auto-connects with your saved token. If it's open in several files at once, the newest connection wins and the others go idle (no flapping).

• Image not shown in Copilot — use a vision-capable model. yfigma_get_screenshot also accepts saveToFile: true to write the PNG into the Codespace (still no S3).

Contributing

Issues and PRs welcome. The bridge/MCP are tiny TypeScript (src/), the plugin is plain JS (plugin/). npm run build must pass.

License

MIT © 2026 Rikuto Yasuda