earmark
MCP server that lets coding agents receive UI feedback from a browser overlay—clicked elements, selectors, source paths, and computed styles—and manage it with watch, ask, acknowledge, resolve, and dismiss tools.
README
earmark
Click an element in your running app, say what should change, and your coding agent gets the CSS selector, the source file and line, the component path, the computed styles and the box geometry — instead of "the button on the right looks wrong".
Works in any framework. No build step required for the overlay.
┌─ browser ──────────────┐ ┌─ broker ────────┐ ┌─ agent ─────────┐
│ click → annotate │ POST │ store + SSE │ MCP │ list / watch │
│ pins, panel, markdown │───────▶│ long-poll │◀──────▶│ ask / resolve │
│ │◀───────│ .earmark/*.json │ │ dismiss │
└────────────────────────┘ SSE └─────────────────┘ └─────────────────┘
Try it in 30 seconds
npm install && npm run example
Open http://127.0.0.1:5173/examples/vanilla/, click the arrow in the toolbar
(bottom right) or press alt+a, then click anything on the page.
The landing page and the full documentation are served alongside it at http://127.0.0.1:5173/site/ and http://127.0.0.1:5173/site/docs.html. Both are live at https://nahar-strativ.github.io/earmark/ and each is a single self-contained file with no dependencies.
For live agent sync, run the broker in a second terminal:
npm run server
Install
npm install -D earmark
import { createEarmark } from 'earmark';
if (import.meta.env.DEV) {
createEarmark();
}
No bundler:
<script type="module" src="/node_modules/earmark/src/index.js" data-earmark-auto></script>
Options
createEarmark({
endpoint: 'http://127.0.0.1:7331', // or false for copy-paste only
hotkey: 'alt+a',
theme: 'auto', // 'auto' | 'light' | 'dark'
persist: true, // keep annotations across reloads
onAnnotate: (annotation) => {},
});
The endpoint defaults to the local broker and degrades silently when nothing is listening — the overlay still works, the sync dot just goes grey.
TypeScript
Every package ships hand-written declarations — there is no build step here, and
an emitted .d.ts would need one. The domain types (Annotation, Target,
Session, Status, Priority) live in earmark and are re-exported by
earmark-server and earmark-mcp, so an annotation is the same type on both
sides of the wire.
import { createEarmark, type Annotation } from 'earmark';
const overlay = createEarmark({ theme: 'dark' });
const pending: Annotation[] = overlay.annotations;
Using it
| Tool | What it does |
|---|---|
| ➤ | Click an element. Shift-click to add more, then click to finish. |
| T | Select text — the exact string is the most greppable thing you can hand an agent. |
| ⛶ | Drag a region. Reports every element inside, or flags an empty area. |
| ❊ | Freeze everything moving — CSS animations, element.animate(), <video>, <audio>. |
| ☰ | Panel: review, delete, answer the agent, copy markdown. |
⌘↵ saves an annotation, esc cancels, alt+a toggles picking. Each
annotation can be marked high, normal or low priority; high sorts
first for the agent.
Copy-paste mode
Click Copy markdown in the panel and paste into your agent:
## UI feedback — 1 annotation
- **Page:** http://localhost:5173/dashboard
- **Viewport:** 1440×900 @2x, dark mode
- **Framework:** react
### 1. Export button padding is too tight — needs 10px 16px
- **Element:** `<button>` <ExportButton>
- **Selector:** `[data-testid="export-btn"]`
- **Source:** `src/components/Card.tsx:42:7`
- **Component path:** App › Dashboard › Card › ExportButton
- **Text:** "Export"
- **Box:** 66×37 at (194, 376)
- **Computed:** padding: 7px 13px; border-radius: 8px; font-size: 13px
- **Ancestors:** div.row ← section.card ← main
Agent sync mode (MCP)
claude mcp add earmark -- npx -y earmark-mcp
Or, to write it into the project's .mcp.json:
npx earmark-mcp init
That one process runs the MCP server and the broker the browser talks to. When something is not working, ask it why:
npx earmark-mcp doctor
✓ Node version: v24.12.0
✓ sqlite backend: available
✓ MCP registration: earmark is registered in .mcp.json
✓ Broker: responding on http://127.0.0.1:7331 — 1 annotations, 2 sessions
✓ Browser overlay: http://localhost:5173/ (1 annotations)
Each failing check prints the command that fixes it, and doctor exits non-zero
so CI can use it.
Tools
| Tool | Purpose |
|---|---|
earmark_list_annotations |
Outstanding work, as markdown (or format: "json"); scope with session |
earmark_watch_annotations |
Blocks until the human annotates something |
earmark_get_annotation |
One annotation with its full reply thread |
earmark_list_sessions |
Which browser tabs are open, and which routes were annotated |
earmark_get_session |
One tab with every annotation it produced |
earmark_acknowledge |
"I've read it, I'm on it" — the pin turns blue |
earmark_ask |
Ask a clarifying question — the pin turns amber |
earmark_resolve |
Mark done with a summary — the pin turns green |
earmark_dismiss |
Decline with a reason the human sees |
earmark_clear |
Delete everything |
earmark_status |
Is the overlay connected? What endpoint should it use? |
The fix loop this enables:
watch → acknowledge → read the source path → edit the file → resolve → watch
acknowledge matters on anything slow: without it, an agent halfway through a
refactor looks exactly like an agent that ignored you. Blue pin means picked up,
green means actually done.
When the feedback is ambiguous, ask instead of guessing. The question appears
on the pin; the human's answer wakes the next watch.
Statuses
open → acknowledged → resolved, with needs-input when the agent is
waiting on a human and dismissed when it declines. Pins are colour-coded:
orange, blue, green, amber, grey.
Sessions
A session is one browser tab, not one page load — the id lives in
sessionStorage, so it survives reloads. Annotations carry their own
page.url, so a session that wandered across three routes gives an agent one
group with three differently-routed items.
SPA navigation is tracked too: pushState, replaceState, popstate and
hashchange all update the session's route list. A tab counts as connected for
exactly as long as its SSE stream is open.
curl http://127.0.0.1:7331/sessions
Source file paths
Selectors tell an agent what to grep for. Source paths tell it exactly where to look, which is the difference between one edit and three greps.
React 19 removed the runtime _debugSource fiber field, so this is done at build
time:
// vite.config.js
import earmark from 'vite-plugin-earmark';
export default {
plugins: [react(), earmark()],
};
Every intrinsic JSX element gets data-earmark-src="src/Card.tsx:42:7" during
vite dev. The plugin also injects the overlay, so createEarmark() in your app
code becomes optional.
earmark({
inject: false, // do not auto-mount the overlay
endpoint: '…', // passed through to createEarmark
applyInBuild: true, // also stamp production builds (off by default)
})
Without the plugin everything still works — you get selectors, component names
and text, just not file:line. You can also add data-earmark-src by hand.
Next.js
Next compiles with SWC, so a Babel plugin would silently switch the whole project
off SWC and slow every build down. earmark-loader is a pre-loader instead — it
sees the source you wrote and leaves the rest of the pipeline alone. It covers
both webpack and Turbopack:
// next.config.mjs
import { withEarmark } from 'earmark-loader/next';
export default withEarmark({
// your config
});
withEarmark(config, {
applyInBuild: true, // stamp production builds too (off by default)
root: process.cwd(), // project root for the reported paths
exclude: 'legacy/', // regexp source string
})
Both compilations are stamped, client and server. That is deliberate: Next renders your components on the server, and React hydration will not add an attribute the server HTML did not have — stamping one side only means the attribute goes missing until something re-renders, with a hydration mismatch on the way.
The wrapper does not mount the overlay for you. Next has no index.html to
inject into, so add it once in a client component:
'use client';
import { useEffect } from 'react';
export function Earmark() {
useEffect(() => {
if (process.env.NODE_ENV !== 'development') return;
import('earmark').then(({ createEarmark }) => createEarmark());
}, []);
return null;
}
Svelte
Vite users — including SvelteKit — get .svelte stamping from
vite-plugin-earmark with no extra configuration. Every plain element in the
markup is stamped before the Svelte compiler sees it:
<button data-earmark-src="src/lib/Card.svelte:12:3" class="go">Go</button>
Components, <svelte:*>, <slot>, comments, {expressions} and the contents of
<script>/<style> are left alone — a stamp in the wrong place there would not
produce a wrong line number, it would produce a file that does not compile.
For a Svelte build that is not Vite, use the preprocessor:
// svelte.config.js
import { earmarkPreprocess } from 'earmark-stamp';
export default { preprocess: [earmarkPreprocess()] };
Svelte still has no runtime hook for component names, so stamping is the whole story there — but CSS rule mapping (below) works regardless.
Plain HTML and CSS — no build step
A static site has no build to stamp, so earmark resolves the source at annotation time instead:
- HTML — the document is re-fetched and parsed with position tracking, then the element's child-index path is walked in the source. Every step is checked against the live tag name, so a framework-rendered page (where the served HTML is just a shell) reports nothing rather than inventing a line.
- CSS — every rule that matches the element, mapped back to the file and line that declares it. This one works everywhere, framework or not.
- **Source:** `index.html:101:11` _(resolved from the served HTML)_
- **CSS rules that style it:**
- `button` → `index.html (inline <style>):49`
- padding: 7px 13px; border-radius: 8px; border: 1px solid var(--line);
- `button.primary` → `index.html (inline <style>):59`
- background: var(--accent); color: rgb(255, 255, 255);
The agent now knows the padding it has to change lives at line 49 in the generic
button rule, not in .primary. Inline <style> blocks are offset into their
host document; external stylesheets report their own path; cross-origin
stylesheets are skipped because their contents are unreadable.
Standalone broker
npx earmark-server --port 7331
curl http://127.0.0.1:7331/markdown
| Route | |
|---|---|
GET /health |
liveness + counts |
GET /annotations?status=open&session=ID |
list |
POST /annotations |
create (batch) |
GET /annotations/wait?since=N&timeout=30000 |
long-poll |
PATCH /annotations/:id |
update status |
POST /annotations/:id/replies |
append to the thread |
DELETE /annotations/:id · DELETE /annotations |
remove · clear |
POST /session |
register a tab / record a route change |
GET /sessions · GET /sessions/:id |
tabs, with counts and annotations |
GET /events?session=ID |
SSE stream; also the tab's liveness signal |
GET /markdown |
the agent-facing document |
Flags: --host --store --file --no-persist --webhook --token --quiet.
Storage
--store json (default) writes a readable .earmark/annotations.json on a
250 ms debounce. --store sqlite writes each change immediately to
.earmark/annotations.db through node:sqlite, so a crash loses at most the
statement in flight — no dependency, Node 22.5+, and it falls back to JSON if
unavailable. --store memory keeps nothing.
Webhooks
npx earmark-server --webhook https://hooks.example/earmark
Also EARMARK_WEBHOOK_URL and EARMARK_WEBHOOKS (comma-separated). Every
annotation event is POSTed with an x-earmark-event header. Delivery is
fire-and-forget with a 5 s timeout and one retry, so a dead endpoint cannot
stall the annotation loop.
Security
This is a development tool.
- The broker binds
127.0.0.1only. Do not bind it to0.0.0.0. - CORS is open by design — your dev server is on an arbitrary origin.
- Any page open in your browser can reach a loopback port. Pass
--token SECRETif that matters on your machine. - Webhooks send annotation content off your machine — page URLs, element text, and whatever you typed. Only configure endpoints you control.
- Source resolution re-fetches your own page and stylesheets from the same origin. Nothing is sent anywhere.
- Do not run it on a shared or public host.
Tests
npm test
Eleven suites, 122 tests: store and HTTP behaviour, the MCP surface driven by a
real stdio client, the overlay's sync client, both persistence backends, webhook
delivery, the init/doctor CLI, the source resolvers, JSX and Svelte stamping,
the webpack/Turbopack loader, and a type-level check of the published
declarations.
npm run types
Type-checks the hand-written declarations. To check the assembled product rather than its units:
npm run verify
Twenty features against real servers, a real MCP process over stdio, real files and a real webhook listener: every broker route, all three storage backends across a restart, the eleven MCP tools, stamping through all three integrations, and both CLI commands. It exists because a bug once passed 122 unit tests and failed the moment the pieces were put together.
Type-checks test/types/check.ts against the hand-written .d.ts files. The
@ts-expect-error lines in it are assertions too — they fail the build if the
error they name stops happening.
Not supported
No screenshots, by decision rather than omission. Your agent already drives a
browser; earmark hands it a verified selector and a URL, so it can capture the
element itself at full fidelity. Shipping html2canvas would send a re-render of
your page rather than what the browser painted, which is the wrong answer for a
tool whose promise is "here is what I am looking at".
Closed shadow roots are opaque to every script, so a component using one is annotated as the component; open roots are picked into and reported with the expression that reaches the element. Nothing inside a shadow root or a canvas has a source line, because that markup is script-created.
Cross-origin iframes stay invisible, because that is a browser security boundary rather than a gap to work around; same-origin frames are pickable, and their own HTML and CSS are resolved. A canvas has no DOM, so earmark reports its coordinate space rather than pretending otherwise. Touch works, but this is still a tool for the machine you develop on. See plan.md for the full open list and the reasoning behind every design decision.
Releasing
All six packages are on npm. To cut the next version, bump the versions (keeping the cross-package pins in step, or a fresh install pulls the older ones underneath) and then:
npm run release
That runs the suite and the live verification first, then
npm publish --workspaces, which publishes in dependency order. It needs
npm login, and publishing cannot be undone, so it is a command you run rather
than one that runs for you.
npm pack --dry-run inside any package shows exactly what would ship. It proves
the file list but not the package: an exports map can still block a subpath that
tools expect, which is only visible once something installs and resolves it.
License
MIT. Clean-room implementation — not derived from any other tool's source.
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。