Figma Agent

Workflow-driven Figma skill for OpenClaw — inspect designs directly, then create, review, and edit screens through structured Figma workflows.
Built on the official Figma Remote MCP server.

What this skill does • Core architecture • Supported workflows • Quick Start • Image Delivery • Limitations

---

What this skill does

Figma Agent is not just a raw MCP wrapper. It is a structured OpenClaw skill for working with Figma in two modes:

• Direct read / inspect for screenshots, metadata, design context, variables, and Code Connect inspection
• Structured write / edit workflows for building screens, reviewing existing work, applying tokens, and iterating on designs through ACP coding sessions

Typical use cases:

• "Show me this screen and tell me what is off"
• "Build the next step based on this existing screen"
• "Replace hardcoded colors with design-system tokens"
• "Inspect local variables and Code Connect mappings"
• "Create a production-ready screen using existing components"

---

Core architecture

The skill is intentionally split into four layers:

1. SKILL.md — routing surface for deciding direct read vs ACP write/edit
2. references/workflow-selection.md — workflow and strategy selection
3. references/core-rules.md — global execution rules and constraints
4. references/prompting-patterns.md + playbooks — prompt structure and step-by-step execution

Runtime model

• Read / inspect goes directly to Figma MCP
• Write / create / edit goes through an ACP coding session with Figma MCP available

This hybrid model keeps inspection fast while preserving a stronger execution path for canvas changes.

Execution model

The controller should operate in three phases:

1. Route & Brief — choose the workflow, identify the actual risks, and produce a lean execution brief
2. Execute — perform the direct read or ACP write/edit task
3. Done Gate — require structural checks first and screenshot confirmation second before reporting success

---

Supported workflows

Read / Inspect

• Read-only inspection — screenshots, metadata, design context, variables, Code Connect context
• Variable Discovery — inspect local tokens and styles before searching externally
• Design Audit Review — review a finished design and identify issues or cleanup opportunities

Create / Edit

• Native Screen Generation — production-ready, design-system-aligned screen creation
• Screen Review Loop — screenshot, isolate issues, apply targeted fixes, validate
• Color Tokenization — replace hardcoded values with variable bindings
• HTML-to-Figma Prototyping — rapid exploration from HTML/CSS, then cleanup
• Stitch Import Cleanup — fix imported Stitch output for native Figma quality
• State Variants / Next Steps — prefer Copy + Edit over rebuild when most of the screen already exists

Execution principles

• Design-system-first — use local variables, styles, components, and Code Connect before creating anything raw
• Component-instance-first — if a suitable component exists, instantiate it instead of rebuilding it visually
• Section-by-section — build in small validated steps, not giant write calls
• Validate after every write — structural checks first, screenshots second

---

Quick Start

Prerequisites

• OpenClaw installed
• Node.js available
• Figma account connected to a supported MCP client
• Full Seat in Figma for write operations (use_figma)

Install

Via ClawHub (recommended):

clawhub install figma-agent

Manual install:

cd ~/.openclaw/skills
git clone https://github.com/rasimme/figma-agent.git
cd figma-agent
npm install

Bootstrap authentication

node ~/.openclaw/skills/figma-agent/scripts/bootstrap-token.mjs

The bootstrap script extracts or refreshes the Figma MCP token from a supported MCP client and writes the Figma server config into your OpenClaw config.

Then restart OpenClaw:

openclaw gateway restart

Verify

Check that Figma MCP is configured in ~/.openclaw/openclaw.json under mcp.servers.figma, then try a simple read action such as get_screenshot or get_metadata.

---

Image Delivery

For chat delivery, screenshots should be sent as actual image attachments, not inline base64 blobs.

Recommended pattern

1. Save the screenshot to disk with --out
2. Reply with MEDIA:<path>

Example:

node scripts/figma-mcp-cli.mjs get_screenshot \
  fileKey=<fileKey> nodeId=<nodeId> scale=2 \
  --out ~/workspace-dev-botti/screenshots/validate.png

Then in chat:

MEDIA:screenshots/validate.png

Important distinction

• image tool = image analysis / vision input
• MEDIA:<path> = actual chat attachment delivery

This matters especially on Telegram, where inline base64 is not the same as a native image attachment.

---

Tool coverage

The skill supports the official Figma Remote MCP toolset for inspection and canvas operations. Commonly used tools include:

• get_design_context
• get_screenshot
• get_metadata
• get_variable_defs
• search_design_system
• get_code_connect_map
• get_figjam
• whoami
• use_figma
• create_new_file
• generate_figma_design
• generate_diagram
• Code Connect mapping helpers

For full tool-level details, see references/figma-api.md.

---

Limitations

Current known constraints:

• Write operations require a Full Seat in Figma
• Large write calls are fragile — split work into smaller sections
• Remote MCP write behavior still has edge cases — validate after each change
• Some Plugin API behavior differs in Remote MCP — see references/plugin-api-gotchas.md
• Image delivery works best through local file output + MEDIA:
• State variants should usually duplicate and edit existing screens rather than rebuild from scratch
• Playbooks guide the agent, not ACP automatically — critical execution constraints must still be injected into ACP prompts when relevant

---

Project structure

figma-agent/
├── SKILL.md
├── README.md
├── CHANGELOG.md
├── LICENSE
├── package.json
├── references/
│   ├── core-rules.md
│   ├── figma-api.md
│   ├── plugin-api-gotchas.md
│   ├── prompting-patterns.md
│   └── workflow-selection.md
└── scripts/
    ├── bootstrap-token.mjs
    ├── figma-mcp-cli.mjs
    ├── figma-mcp.mjs
    └── token-scanner.mjs

---

License

MIT — see LICENSE.