loupe-mcp
Turns product feedback pinned to a live UI into an actionable backlog for Claude Code — list comments, open one with its target element's HTML, computed styles and screenshot, and update its status.
README
Loupe
Pin feedback to the live UI. Hand it to Claude.
Loupe lets product managers inspect any element on a running product, pin a comment to it, and capture a screenshot — then hand that feedback straight to Claude Code as an actionable, fully-contextualized backlog. Comments persist and re-anchor across redeploys.
📖 Read the full documentation →
This is a monorepo (npm workspaces). Every piece runs locally with no external
services — the database is embedded Postgres (PGlite), so npm install && npm run build && npm run seed && npm start is the whole setup.
Packages
packages/
shared/ canonical types + normalizeUrl() (built to dist, consumed by all)
sdk/ embeddable browser SDK — inspect, comment (element / region / free note), capture, re-anchor; draggable toolbar
demo/ a fake product ("Acme Analytics") to try it on
server/ comment API — node:http, Postgres, object storage, HMAC auth, static hosting
dashboard/ Kanban triage board (reads the API)
mcp/ MCP server — exposes comments to Claude Code
extension/ MV3 browser extension — inspect/comment on ANY site, pixel-perfect capture
laravel/ Loupe for Laravel — composer package: widget + your DB + gating + dashboard + MCP
Loupe for Laravel
Prefer to run the whole loop inside your own app? loupekit/laravel
is a Composer package that embeds the widget, stores comments in your database, gates
access with your authorization rules, serves the dashboard on your routes, and
exposes the backlog to Claude over MCP — no separate Node backend. See the
full guide.
composer require loupekit/laravel
php artisan loupe:install && php artisan migrate
# add @loupeWidget to your layout, then open /loupe/dashboard
Run it
npm install
npm run build # shared → sdk → dashboard → extension
npm run seed # creates the demo project; prints its admin key + demo HMAC
npm start # one process on http://localhost:8787 serves API + dashboard + demo + SDK
Then open:
- Demo product — http://localhost:8787/demo/ (leave feedback; persists to the API)
- Triage board — http://localhost:8787/dashboard/?key=<admin key from
npm run seed>
Point Claude Code at the comments:
{
"mcpServers": {
"loupe": {
"command": "node",
"args": ["/absolute/path/to/loupe/packages/mcp/index.ts"],
"env": {
"LOUPE_API": "http://localhost:8787",
"LOUPE_PROJECT_KEY": "pk_demo_acme",
"LOUPE_ADMIN_KEY": "<admin key from npm run seed>"
}
}
}
}
Then: "list the open Loupe comments and work through them." Claude calls
list_comments → get_comment (request + element HTML + computed styles + screenshot
URL) → makes the change → update_status.
Architecture notes
Database — server/db.ts is a one-function query() seam. With DATABASE_URL set
it uses node-postgres against real/hosted Postgres; otherwise embedded PGlite on disk.
Same SQL, same $1 params.
Auth — every project has a secret. Writes require X-Loupe-User +
X-Loupe-Hmac = HMAC-SHA256(userId, secret) (the host app's server computes this and
injects it — the demo's value is precomputed by npm run seed). The dashboard and MCP
server authenticate as admin with X-Loupe-Admin = secret. Screenshot blobs are served
by unguessable id (prod: signed URLs).
Object storage — server/blobs.ts is the seam (local disk now, S3 later). The SDK
uploads a screenshot to POST /v1/blobs and stores only the returned URL on the
comment, so lists/reads stay small.
URL normalization — normalizeUrl() in @loupekit/shared strips utm_*, click ids,
and Loupe's dev params, so a comment on /checkout?utm_source=x and one on /checkout
don't fragment. Applied server-side on write and query.
Browser extension
The extension reuses the exact SDK core; its only difference is the screenshot source —
chrome.tabs.captureVisibleTab (real pixels), cropped to the element with redaction,
wired via the SDK's captureScreenshot override. Load it manually:
npm run build:extension # builds packages/extension/content.js
chrome://extensions→ enable Developer mode → Load unpacked → selectpackages/extension.- Click the Loupe icon → set project key (
pk_demo_acme), user, API base (http://localhost:8787), and (optionally) the demo HMAC → Start Loupe on this tab.
Note: the extension is validated by build + manifest/bundle checks; full in-browser E2E wasn't automated in this repo's headless setup.
Roadmap
- Client SDK ✓ — inspect, comment, capture, re-anchor across redeploys.
- Backend API ✓ — Postgres, HMAC auth + per-project secrets, object storage, static hosting.
- MCP server ✓ — comments as a Claude Code backlog.
- Dashboard ✓ — Kanban triage.
- Hardening ✓ — monorepo, Postgres, enforced auth, blob storage, URL normalization.
- Browser extension ✓ — pixel-perfect capture on any site.
Next: real cloud object storage (S3/R2) + signed URLs, project/team management UI, Postgres migrations tooling, and packaging the SDK/MCP to npm + the extension to the Chrome Web Store.
Author
Loupe is created and maintained by Mohamed Ashraf Elsaed.
- 💼 LinkedIn: mohamedashrafelsaed
- 🐙 GitHub: @mohamed-ashraf-elsaed
- ✉️ Email: m.ashraf.saed@gmail.com
If Loupe is useful to you, a ⭐ on the repo and a connection on LinkedIn are always appreciated. For consulting or collaboration, reach out via any of the links above.
License
MIT © Mohamed Ashraf Elsaed
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。