Spec Driven Development
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
作者提供的说明(英文原文)
Author, validate, lay out, render, and diff architecture diagrams as code.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config,把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"arc": {
"command": "npx",
"args": [
"-y",
"@hudsonkit/arc-mcp"
]
}
}
}在终端中运行这条命令
claude mcp add arc -- npx -y @hudsonkit/arc-mcp粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"arc": {
"command": "npx",
"args": [
"-y",
"@hudsonkit/arc-mcp"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"arc": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@hudsonkit/arc-mcp"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月7日
注册名称已通过 GitHub 账户 hudsonkit 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(0.14.1)于12天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。
同一分类中有:Spec Driven Development, Remote Desktop Commander, Tools MCP
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
Remote Desktop Commander: use files, terminal and processes on your own computers from any AI app.
Web production tools for AI agents: contrast, JP text, WebP, JSON-LD, slugs, dates and 40+ more
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
Gives AI agents the full Arc toolchain over the Model Context Protocol — author, validate, lay out, render, and diff architecture diagrams from any MCP client.
No install needed — the server runs over stdio:
npx -y @hudsonkit/arc-mcp
Requires Node.js 22 or newer. No account, API key, or environment configuration is required. PNG export needs a local Chrome/Chromium installation; animation export also needs ffmpeg.
Claude Code:
claude mcp add --scope project arc -- npx -y @hudsonkit/arc-mcp
Devin:
devin mcp add arc -- npx -y @hudsonkit/arc-mcp
Any MCP client (JSON config):
{
"mcpServers": { "arc": { "command": "npx", "args": ["-y", "@hudsonkit/arc-mcp"] } } }
| Tool | What it does |
|---|---|
validate_diagram | Coded diagnostics for ArcDiagramData — consume by code, apply a supportedFixes entry, re-validate |
auto_layout | Auto-layout a full diagram or minimal input (nodeData + connectors + connectorStyles) |
diff_diagram | Structural DiagramDelta — feed to for a diff render |
render_svg | Deterministic SVG export — eleven themes, light/dark; authored flows animate via SMIL or sample at flowTime |
render_png | PNG raster via a local Chrome/Chromium binary (ARC_CHROME to override, ARC_CHROME=none to disable), sampled at flowTime |
render_animation | GIF/MP4 flow animation to an output path via Chrome frames + ffmpeg (ARC_FFMPEG to override) |
render_html | React component snippet, iframe embed, or standalone HTML page |
render_ascii | Unicode/ASCII box-art rendering for terminals and docs |
diagram_to_typescript | Emit a typed ArcDiagramData TypeScript module |
editor_handoff | Build URLs that open the diagram in the Arc studio |
open_live_diagram | Open a live architecture view and return its stable URL and session ID |
begin_live_diagram_work | Declare work before starting it, with a summary and canvas, node, or region scope |
update_live_diagram | Publish a complete validated diagram revision to that view |
finish_live_diagram_work | Finish the current work ID and leave the view idle for the next task |
end_live_diagram | Mark the diagram complete and retain the final view |
open_architecture_review | Save a baseline and open a durable architecture review |
update_architecture_review | Save findings, evidence, decisions, questions, and summary |
resume_architecture_review | Resume an interrupted review with its author capability |
revoke_architecture_review_share | Invalidate the old viewer link and return a new capability |
Start the Arc studio with bun run dev, then call open_live_diagram with { diagram, title?, agent?: { harness, name? } }. Open the returned liveUrl once. The view starts idle. Before doing diagram work, call begin_live_diagram_work with { sessionId, summary, scope? } and retain its workId. Publish complete revisions with update_live_diagram({ sessionId, diagram, message? }); keep node IDs stable so the viewer can follow each move. In your cleanup/finally path, call finish_live_diagram_work({ sessionId, workId, message? }) when the declared work ends. This leaves the same view idle and reusable. Use end_live_diagram with { sessionId, message? } only when the entire diagram session is complete.
Scope defaults to { type: "canvas" }. Use { type: "nodes", nodeIds: ["api"] } for existing boxes, or { type: "region", x: 400, y: 80, width: 200, height: 160 } for the area where new boxes will appear. A scope accepts 1–200 unique existing node IDs, each at most 200 characters. Region coordinates must be finite and within ±100000; width and height must be positive and at most 100000. Summaries are trimmed, 1–500 characters, and cannot contain control characters. Concurrent begin calls and stale finish IDs are rejected without replacing active work.
The live view belongs to the Hudson Kit Arc studio. ARC_EDITOR_URL selects the studio base URL; open_live_diagram also accepts baseUrl for an individual view, including a deployment prefix such as /arc. The MCP client name/version comes from its initialize handshake. The displayed connection ID is generated by the server; it does not claim to be the client's conversation ID.
The optional agent object declares a harness label and, if useful, an agent name for this view. For example, agent: { harness: "Codex", name: "Reviewer" } adds actor.harness and actor.displayName. These are caller-declared labels, not verified identities. They never replace the actual MCP actor.name, actor.version, or connection actor.id, and do not carry over to other views. Both labels are trimmed, limited to 1–80 characters, and reject control characters.
The local scripts/liveDiagram.control.ts driver accepts ARC_AGENT_HARNESS and optional ARC_AGENT_NAME to send these labels. Neither has a default; ARC_AGENT_NAME requires ARC_AGENT_HARNESS. The driver accepts JSON lines such as {"type":"begin","summary":"Inspect API","scope":{"type":"nodes","nodeIds":["api"]}}, then {"type":"finish","workId":" ","message":"API checked"}. Its status command reports local session metadata. quit closes the relay; it does not claim that unfinished work succeeded.
In stdio mode the first live view starts a loopback HTTP relay on a free port. Keep the MCP process running and open the view on the same computer. The browser may request local-network or loopback-network access when a hosted HTTPS studio connects to that relay. Allow that permission for the studio. Loopback addresses are treated specially by browser mixed-content rules; arbitrary HTTP LAN or remote relay URLs do not get the same guarantee. Browser support, enterprise policy, and the studio's CSP can still prevent the connection. An all-local studio (bun run dev, ARC_EDITOR_URL=http://localhost:5188) avoids the public website-to-loopback permission boundary. For a remote relay used from an HTTPS studio, serve the relay over HTTPS. See the Local Network Access specification and Chrome's permission guidance. HTTP mode serves the stream beside /mcp, so a configured remote relay can be used from another computer. The URL fragment carries a read capability for one diagram; anyone given the full URL can read it. Only its owning MCP connection can update it. Browser requests must come from that view's configured studio origin. The HTTP server enforces ARC_MCP_TOKEN on every MCP request when set, and requires it when binding outside loopback.
Each stream sends the latest full snapshot immediately, including after a reconnect. Invalid updates preserve the previous state. Only explicit work is working; edits outside a declared work item stay idle. Valid updates during work preserve its scope and renew a five-minute lease. Finishing marks the work complete and the session idle. Disconnect or lease expiry marks work interrupted; ending the entire session completes any active work. A later standalone edit clears previous work metadata. These calls describe declared diagram work, not unobserved harness thinking. Unfinished diagrams remain available while their owning MCP connection stays open. Completed diagrams expire one hour after completion; unfinished diagrams expire one hour after their owner disconnects. Stopping the MCP process discards ordinary in-memory live sessions. Save/export a diagram from the studio when it needs to outlive that process.
Live tools require a matching MCP build and studio build. During development, build this checkout with bun run build:mcp and configure the client to run node /absolute/path/to/arc/packages/mcp/dist/arc-mcp.mjs; run its studio with bun run dev. Listing a tool here does not mean an older published package or the currently deployed Hudson Kit website already includes it.
The private preview adds a durable review to the same Hudson Kit canvas. Follow the private preview setup to build matching studio and MCP artifacts and serve them from one private address.
Open with open_architecture_review({diagram,title,question,repository?,agent?}). Repository metadata is {name,url?,revision?}. The result includes a private author resumeToken as well as the normal sessionId and read-only viewer URL. Use the existing begin/update/finish/end tools for declared work and diagrams. Save evidence through update_architecture_review with summary, findings, decisions, and openQuestions; omitted fields stay unchanged, and supplied arrays replace the full previous list.
Findings are {id,title,detail,severity,nodeIds?,sources}, with severity info|low|medium|high. Sources are {path,line?,revision?,url?}. The studio shows those references alongside the canvas, and its Save review action downloads {schemaVersion:1,baseline,snapshot,revisions} without secrets. Source claims are author-supplied evidence, not verified by Arc itself.
Reviews persist before publishing each revision. Keep ARC_REVIEW_DIR and the private server address stable to reopen the same saved viewer link after a restart. Use resume_architecture_review({sessionId,resumeToken,agent?}) after losing the author connection. Optional agent: {harness,name?} supplies caller-declared display labels alongside the new connection's actual MCP identity. Omit it to use only that connection identity. It cannot displace a live author. Completed reviews stay read-only and keep their final author and history unchanged. revoke_architecture_review_share({sessionId}) invalidates the old viewer capability. Saved reviews are not removed by the ordinary live view's one-hour expiry.
For a terminal author, scripts/review.control.ts keeps one HTTP MCP connection open and accepts JSON lines of {id?,name,arguments}. Set ARC_MCP_URL and, if configured, ARC_MCP_TOKEN in its environment. EOF disconnects the author; finish/end calls must be explicit. Keep its output private because opening a review returns author and viewer capabilities.
| URI | Content |
|---|---|
arc://schema/diagram | Generated draft-07 JSON Schema for ArcDiagramData |
arc://skill/diagrams | Arc diagram-generation skill for agents |
arc://docs/llm | Dense LLM briefing |
The published bin runs over stdio, which is all a local client needs. To serve Arc to clients on other machines, run the Streamable HTTP server from a checkout of the repo:
git clone https://github.com/hudsonkit/arc && cd arc
bun install ARC_MCP_HOST=127.0.0.1 ARC_MCP_PORT=5190 bun run mcp:http
| Variable | Default | Purpose |
|---|---|---|
ARC_MCP_PORT | 5190 | Port to listen on |
ARC_MCP_HOST | 127.0.0.1 | Interface to bind; nonloopback binds require ARC_MCP_TOKEN |
ARC_MCP_PUBLIC_URL | http:// : | Externally reachable origin used in live stream URLs and Host validation; preserve this Host through a reverse proxy |
| ARC_MCP_ALLOWED_ORIGINS | Studio and relay or
只显示了开头部分;其余内容请在仓库中阅读