anriss
Enables UI feedback loop by clicking elements, leaving comments, and letting AI coding agents (via MCP) resolve annotations interactively.
README
<p align="center"><img src="docs/logo.svg" width="88" alt="Anriss logo"></p>
Anriss
Click elements on your running web app, leave a comment, and let your AI coding agent fix it. A tiny self-hosted feedback loop for UI work — no React, no browser extension, no SaaS.
Anriss — from German anreißen: marking out a workpiece before machining. And Riss is the plan itself (Grundriss, Aufriss). Mark it. Plan it. Fix it.
you (browser) anriss your AI agent
────────────── ───────────── ─────────────
click elements ──POST──▶ Rust + SQLite ◀──MCP── Claude Code /
+ one comment W3C annotations Codex / opencode
fixes code,
marks resolved
- One script tag to integrate — works with any frontend (plain HTML, Stimulus/Turbo, Vue, React, server-rendered, anything)
- Multi-select: click several elements, leave one comment for all of them
- Agent-native: built-in MCP endpoint; any MCP-capable coding agent can list open annotations and resolve them
- Standards-based: annotations are stored as W3C Web Annotation Data Model JSON (CSS selector + XPath + HTML snippet per target)
- Self-contained: single Rust binary with the widget embedded; SQLite file as the only state
⚠️ Development tool. CORS is wide open by design and there is no authentication. Run it on localhost or inside your dev network only.
Quick start
It's a single process — just run the binary:
cargo install --path . # once; installs `anriss` into ~/.cargo/bin
anriss # serves on :4855, SQLite in ./data/annotations.db
Or straight from the checkout:
cargo run --release
Open http://localhost:4855/ — the built-in demo page runs the widget, so you can try the whole loop before touching your own app.
Building from a fresh clone only needs Rust; the widget bundle
(dist/anriss.js) is committed. Rebuild it after widget changes with
npm install && npm run build.
A Dockerfile is included if you prefer a container:
docker build -t anriss . && docker run -p 4855:4855 -v "$PWD/data:/data" anriss
Add the widget to your app (dev only!)
<script src="http://localhost:4855/anriss.js" defer></script>
Gate it however your stack gates dev-only markup, e.g. Symfony/Twig:
{% if app.debug %}<script src="http://localhost:4855/anriss.js" defer></script>{% endif %}
Annotate
- Click the Annotate pill (bottom-left) or press Alt+Shift+A
- Click one or more elements — click again to deselect
- Press Enter or click Comment, write what should change, save
Connect your agent (MCP)
Claude Code — .mcp.json in your project:
{
"mcpServers": {
"annotations": { "type": "http", "url": "http://localhost:4855/mcp" }
}
}
Codex — ~/.codex/config.toml:
[mcp_servers.annotations]
url = "http://localhost:4855/mcp"
opencode — opencode.json:
{
"mcp": {
"annotations": { "type": "remote", "url": "http://localhost:4855/mcp" }
}
}
Then just ask: "Work through the open UI annotations." The agent calls
list_open_annotations, locates each element via the stored selectors and
HTML snippets, fixes the code, and calls resolve_annotation.
Lifecycle & live updates
open ──agent resolves (MCP/REST)──▶ done ──you tick ✓ in the widget──▶ archived
- The widget polls the server every 3 s (paused while the tab is hidden), so agent resolutions appear live in the pill's list panel — resolution note included — without a reload. A red dot on the pill means the server is unreachable.
- Resolve never deletes. Archiving is deliberately human-only (no MCP tool for it): you review the fix, then tick it away.
- The panel's Copy JSON button copies all open annotations — the manual fallback when you want to paste them into any agent directly.
REST API
| Method | Path | Description |
|---|---|---|
GET |
/anriss.js |
The drop-in widget |
POST |
/api/annotations |
Create an annotation (the widget calls this) |
GET |
/api/annotations?status=open|done |
List annotations |
POST |
/api/annotations/{id}/resolve |
Mark resolved: {"resolution": "..."} |
GET |
/healthz |
Health check |
Annotations are W3C Web Annotations with ext:* extension fields for
status/resolution. Comments may reference their targets as [#n] — n is the
1-based index into the annotation's target[] array (the widget's numbered
chips/badges use the same numbering). Each target carries a CssSelector, an XPathSelector
and a truncated HTML snippet, so agents (and future UI pins) can re-anchor
even after the DOM changed.
Configuration
| Env var | Default | |
|---|---|---|
ANRISS_PORT |
4855 |
HTTP port |
ANRISS_DB |
./data/annotations.db |
SQLite path |
Widget: <script ... data-hotkey="ctrl+shift+e"> overrides the hotkey.
Development
npm run build # compile Tailwind subset + inline into dist/anriss.js
cargo test # server tests
cargo clippy
Queries are compile-time checked (sqlx). After changing SQL or migrations:
DATABASE_URL=sqlite://.sqlx-dev.db cargo sqlx migrate run && cargo sqlx prepare
(the .sqlx/ offline cache is committed, so plain cargo build needs no DB).
dist/anriss.js is committed so the crate builds without Node.
The widget UI lives in a Shadow DOM and its Tailwind classes are compiled
with important — host-page styles can't break it, and it can't leak out.
Marketing page
docs/index.html is both the landing page and the local demo/test bed:
- Served by the binary at
/(embedded at compile time) — there the widget is live and annotations work end to end. - Publishable via GitHub Pages: repo Settings → Pages → Build and
deployment → Source "Deploy from a branch" → branch
main, folder/docs. On Pages the widget script simply 404s (no server) — the page itself works.
License
MIT
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。