widget-studio
An agent capability for creating and displaying rich UI widgets (cards, dashboards, task lists, charts, forms, tables) inline in conversations — packaged as a Claude Code plugin, an Agent Skill, standalone CLI scripts, and an MCP server with MCP Apps support. Widgets use the ChatKit widget JSON format, generated by prompting Widget Studio's hosted generator; the JSON is edited directly only for small tweaks.
Architecture
The MCP server is optional — the skills + CLI deliver the full generate→render loop on their own. The server is the layer that adds interactive inline widgets in MCP Apps hosts. Each host gets the richest rendering it supports:
| Layer | What it provides | Who uses it |
|---|---|---|
| Skills (skills/) | /widget-studio:make — how to prompt the generator well; :tweak — instant JSON edits + re-display; :preview — render saved JSON to PNG/HTML; :doctor — install/render/API health checks | Any skills-capable agent |
| Reference (skills/make/references/chatkit-widgets.md) | ChatKit component schema for reading and making small edits to generated widgets before re-display | The model, on demand |
| Scripts (scripts/, bin/) | widget-generate (prompt → widget via hosted API), widget-render (JSON → PNG / standalone HTML), widget-export (→ .widget file for the Widget Studio editor) — self-locating, on PATH in plugin sessions | The full capability, no MCP needed |
| MCP server (src/) | generate_widget, display_widget, and export_widget (.widget files) tools; interactive MCP Apps iframe with a copy/export action bar in capable hosts, automatic PNG fallback elsewhere | MCP Apps hosts (interactive) + any other MCP client (PNG) |
| Evals (evals/) | Deterministic golden-render suite (npm run eval) + skill-creator-format LLM eval cases (skills/make/evals/evals.json) | CI / skill iteration |
Rendering paths
- MCP Apps hosts (Claude Desktop, claude.ai, ChatGPT): interactive iframe via the
standard
_meta.ui.resourceUri→ui://widget-studio/widget.html, a fully self-contained HTML renderer (CSP-safe, light/dark, actions post back to chat). Legacyopenai/outputTemplatealias included for older ChatGPT builds. - Non-UI MCP clients (Claude Code CLI, others): the server detects the missing
MCP Apps capability at initialize and attaches a PNG image of the widget to the
tool result instead (see PNG rendering). Override with
WIDGET_STUDIO_IMAGE=auto|always|never. - No MCP at all:
widget-render(scripts/render-widget.mjs) produces the same PNG or a self-contained HTML preview from a widget JSON file.
Install
Note: the
npxforms and one-click badges below go live with the first npm release ofwidget-studio-mcp. Until then, use the from-source form — everything else is identical.
Install in Cursor Install in VS Code Install in LM Studio
As a Claude Code plugin (recommended for Claude)
/plugin marketplace add ericlewis/widget-studio-mcp
/plugin install widget-studio@widget-studio
The plugin registers the MCP server, puts widget-generate/widget-render/widget-export
on PATH, and installs the skill family (/widget-studio:make, :tweak, :preview,
:doctor). dist/ ships self-contained bundles — no npm install needed.
MCP server, per host
Claude Code
claude mcp add widget-studio -- npx -y widget-studio-mcp@latest
Claude Desktop
claude_desktop_config.json (Settings → Developer → Edit Config):
{ "mcpServers": { "widget-studio": { "command": "npx", "args": ["-y", "widget-studio-mcp@latest"] } } }
Cursor
Use the install badge above, or .cursor/mcp.json:
{ "mcpServers": { "widget-studio": { "command": "npx", "args": ["-y", "widget-studio-mcp@latest"] } } }
VS Code
code --add-mcp '{"name":"widget-studio","command":"npx","args":["-y","widget-studio-mcp@latest"]}'
Codex CLI
codex mcp add widget-studio -- npx -y widget-studio-mcp@latest
Or ~/.codex/config.toml:
[mcp_servers.widget-studio]
command = "npx"
args = ["-y", "widget-studio-mcp@latest"]
Gemini CLI
As an extension (bundles the MCP server and skills):
gemini extensions install https://github.com/ericlewis/widget-studio-mcp
Windsurf
~/.codeium/windsurf/mcp_config.json:
{ "mcpServers": { "widget-studio": { "command": "npx", "args": ["-y", "widget-studio-mcp@latest"] } } }
Goose
goose session --with-extension "npx -y widget-studio-mcp@latest"
Amp
amp mcp add widget-studio -- npx -y widget-studio-mcp@latest
ChatGPT (and other remote-only hosts)
ChatGPT connects to remote MCP servers only. Run the HTTP transport behind a tunnel or deploy, then add the URL under Settings → Connectors (developer mode):
npx -y widget-studio-mcp@latest --http 3400 # stateless streamable HTTP, POST /mcp
ChatGPT implements the MCP Apps standard natively, so widgets render
interactively; in --http mode the server also always attaches the PNG.
Bridge for stdio-only hosts hitting a remote:
npx -y mcp-remote https://your-host/mcp.
From source
git clone https://github.com/ericlewis/widget-studio-mcp && cd widget-studio-mcp
npm install && npm run build
claude mcp add widget-studio -- node "$(pwd)/dist/index.js" # or the equivalent for your host
PNG rendering
Built in — pure JS/WASM (satori + resvg with bundled Inter fonts). No browser, no playwright/puppeteer, no native dependencies, nothing to install.
Develop
npm run build # typecheck + bundle the iframe app and self-contained server into dist/
npm test # protocol smoke test (incl. one real API generation)
npm run eval # render golden fixtures to PNG + offline protocol checks
Manual preview of any widget JSON:
node scripts/render-widget.mjs --in evals/golden/sales-dashboard.json --png dashboard.png
node scripts/render-widget.mjs --in evals/golden/sales-dashboard.json --html dashboard.html
dist/ is committed on purpose: plugin installs copy the repo as-is, so the
bundles must ship prebuilt. Rebuild before committing changes under src/.
Notes
- The Widget Studio API is the unauthenticated hosted service behind the public
editor; requests carry a stable random user ID (
WIDGET_STUDIO_USER_IDto pin). - The renderer covers the full ChatKit component set; unknown components degrade to rendering their children. External images are blocked by host CSP in interactive mode (placeholder shown) but fine in PNG/HTML output.