diagrams-so
The Diagrams.so MCP server enables generating, editing, and managing cloud architecture diagrams from any MCP client, wrapping the public Diagrams.so API.
README
@diagrams-so/mcp
The Diagrams.so MCP server — generate, edit, and manage cloud architecture diagrams from any MCP client (Claude Desktop, Claude Code, Cursor). It's a thin stdio client over the public Diagrams.so API (/api/v2); every tool is one REST call.
Quick start
Needs Node ≥ 18. No API key to copy: you connect by approving in a browser.
# 1. add the server to Claude Code
claude mcp add diagrams-so -- npx -y @diagrams-so/mcp
# 2. connect this machine (a browser opens, press Approve)
npx @diagrams-so/mcp login
Restart your client, then ask: "Generate an AWS 3-tier web app diagram and show me the warnings."
Run /mcp if you want to confirm all 23 tools registered first.
Using it a lot? Install once and the command gets shorter:
npm i -g @diagrams-so/mcp
diagrams-so login
The server talks to production (
https://api.diagrams.so/api/v2) by default. Point it at a local or self-hosted API withDIAGRAMS_API_BASE.
Using Claude Desktop or Cursor instead of the CLI? See Add it to your MCP client. Prefer a one-click, no-terminal install? See One-click install (MCPB).
<details> <summary>No terminal at all (Claude Desktop)</summary>
Add the server to your client config (below), then just ask for a diagram. Because the machine isn't connected yet, the first tool call replies with a link. Open it, press Approve, and ask again. Nothing to install or type. </details>
Generating a diagram from Claude
login only connects the machine. It never generates anything itself, so there is no
generate command to type in a terminal. You write the prompt in your assistant's normal
chat box and it calls the tools for you.
| Client | Where you type the prompt |
|---|---|
| Claude Code | the terminal chat, same place you ask anything else |
| Claude Desktop | the normal message box |
| Cursor | the chat or composer panel |
MCP servers load at startup, so restart your client after adding it. Then run /mcp and
confirm diagrams-so shows 23 tools.
Now just ask, in plain English:
Generate an AWS three-tier web app with an ALB, EC2 Auto Scaling and RDS Multi-AZ. Show me the design warnings, then export it as draw.io.
Behind that one sentence the assistant calls generate_diagram, then get_warnings, then
export_diagram. You never name a tool or write JSON.
More things worth asking, once you have a diagram:
The warnings mention no encryption in transit. Fix that one and show me the new score.
Add a CloudFront distribution in front of the ALB.
Re-export it as SVG so I can drop it in the README.
Each reply carries the credit cost and your remaining balance. Generating, editing, fixing and re-laying-out spend credits; reading, warnings and every export are free.
The .drawio file the assistant saves opens at app.diagrams.net or
in the desktop app, fully editable — it is real draw.io XML, not a picture.
Prefer not to use an assistant at all? diagrams.so/create has
the same thing as a web page: type the prompt in the box. No install, no login, no MCP.
Tools (23)
Create & change (mutating)
| Tool | What it does | Cost |
|---|---|---|
generate_diagram |
Create a diagram from a prompt → id + draw.io XML + warnings + score | credits |
edit_diagram |
Apply a natural-language change (new version) | credits |
fix_warning |
Resolve one Well-Architected warning | credits |
relayout_diagram |
AI re-arrange the layout (async; first 2/diagram free, then confirm) | credits* |
import_diagram |
Import existing draw.io XML as a new diagram | free |
update_diagram |
Rename / change visibility / replace XML | free |
revert_diagram |
Revert to an earlier version | free |
delete_diagram |
Soft-delete a diagram (destructive) | free |
fork_template |
Copy a public/library diagram into your account (private) | free |
Read (free)
| Tool | What it does |
|---|---|
get_diagram |
Fetch a diagram's XML + score |
list_diagrams |
List your diagrams (cursor-paginated) |
get_warnings |
Well-Architected findings for a diagram |
export_diagram |
Raw drawio or svg file (SVG is a paid feature) |
list_versions |
Version history (with is_current) |
get_version |
A specific version's XML + score |
get_relayout_status |
Poll an async re-layout job |
search_gallery |
Search public community + curated library diagrams |
enhance_prompt |
Turn a rough idea into a detailed prompt |
clarify_prompt |
Get clarifying questions for a vague prompt |
get_usage |
Plan + credits + cost estimates |
get_usage_history |
Itemized credit ledger per task (action, credits, diagram, surface) with filters + a live session tally |
whoami |
Account, plan, scopes, live/test mode |
list_capabilities |
Valid diagram types / providers / export formats |
Reads and exports are free; generate / edit / fix / relayout cost credits, and delete is destructive. relayout asks for confirm=true before it charges.
CLI
Commands
Installing the package puts diagrams-so on your PATH. The longer
diagrams-so-mcp name still works, and npx @diagrams-so/mcp <command> works
without installing anything.
npm i -g @diagrams-so/mcp
diagrams-so login # connect this machine (a browser opens, press Approve)
diagrams-so whoami # which account is this machine connected as
diagrams-so logout # remove the local credential
diagrams-so install # print client config to paste
Environment
| Var | Required | Default |
|---|---|---|
DIAGRAMS_API_KEY |
❌ | — (CI and headless only; login is the normal path) |
DIAGRAMS_API_BASE |
❌ | https://api.diagrams.so/api/v2 (production; override for local/self-hosted) |
DIAGRAMS_API_TIMEOUT_MS |
❌ | 450000 (per-request safety-net; sits above the API's full server-side timeout ladder) |
DIAGRAMS_BROWSER |
❌ | — (login opens your default browser; set e.g. Google Chrome when your Diagrams.so session lives in a non-default browser) |
DIAGRAMS_NO_BROWSER |
❌ | — (set to any value to stop login opening a browser; the URL is always printed) |
DIAGRAMS_NO_AUTO_LOGIN |
❌ | — (set to any value to disable in-tool connect; unauthenticated tools then just say to run login) |
Add it to your MCP client
The published package runs straight from npm, so there is no path to fill in and no key in the
config. Run npx @diagrams-so/mcp install to print these blocks for your client.
Claude Code (CLI)
claude mcp add diagrams-so -- npx -y @diagrams-so/mcp
npx @diagrams-so/mcp login
Claude Desktop / Cursor (claude_desktop_config.json / mcp.json)
{
"mcpServers": {
"diagrams-so": {
"command": "npx",
"args": ["-y", "@diagrams-so/mcp"]
// optional: "env": { "DIAGRAMS_API_BASE": "http://localhost:8000/api/v2" } for a local API
}
}
}
Connect either by running npx @diagrams-so/mcp login once, or by asking for a diagram and
clicking the link the first tool call gives you. Set DIAGRAMS_API_KEY instead only for CI and
headless machines, where no browser can open.
Restart the client, then ask: "Generate an AWS 3-tier web app diagram and show me the warnings."
One-click install (MCPB)
Download diagrams-so.mcpb from the
latest release and drag it into
Claude Desktop. No terminal, and it no longer asks for an API key: install it, then connect on first
use by clicking the link the first tool call gives you.
Building the bundle yourself is a contributor step, see Developing locally.
Verify it works
DIAGRAMS_API_KEY=dgz_live_your_key node test-smoke.mjs
Runs the full flow (connect → list tools → whoami → generate → warnings → fix → export → error handling).
Continuous integration & releases
| Workflow | Trigger | What it does |
|---|---|---|
CI (ci.yml) |
every push / PR | npm ci + npm run build on Node 18/20/22, then a free smoke (scripts/ci-smoke.mjs) that launches the built server and asserts all 23 tools register. No API calls, no credits. |
Live smoke (live-smoke.yml) |
nightly + manual | the full end-to-end flow (test-smoke.mjs) against the real API. Spends credits — runs only when the DIAGRAMS_API_KEY secret is set. |
Release (release.yml) |
tag vX.Y.Z |
build → npm prune --omit=dev → pack diagrams-so.mcpb → attach to a GitHub Release. Publishes to npm too if an NPM_TOKEN secret is set. |
Cut a release:
# bump "version" in package.json + manifest.json to match, commit, then:
git tag v1.2.0 && git push origin v1.2.0
The tag must match package.json's version (the workflow enforces this). The
.mcpb bundle appears on the GitHub Release; manual workflow_dispatch produces
it as a downloadable artifact without publishing (handy for testing a bundle).
Repo secrets/variables (optional): DIAGRAMS_API_KEY (live smoke),
NPM_TOKEN (npm publish), DIAGRAMS_API_BASE variable (non-prod live-smoke target).
Notes
- stdout is the MCP channel — the server logs only to stderr.
- Errors come back as clean MCP tool errors carrying the API's
code, HTTP status, andrequest_id. - The server never talks to internal services or the database — only the public
/api/v2. - Long-running tools stay alive past client timeouts.
generate/edit/fix/relayout/enhance_prompt/clarify_promptemit anotifications/progressevery 10s while running, so MCP clients that reset their request timeout on progress (resetTimeoutOnProgress) won't abort a slow generation at the SDK's 60s default. If your client doesn't reset on progress, raise its per-call timeout for these tools.
Developing locally
Only needed if you are changing the server itself. Users should install from npm, see Quick start.
git clone https://github.com/RedHold/diagrams-mcp-app-core.git
cd diagrams-mcp-app-core
npm install # installs deps and builds dist/ via the prepare hook
node scripts/ci-smoke.mjs # all 23 tools register; no API calls, no credits
# point a client at your working copy
claude mcp add diagrams-so-dev -- node "$(pwd)/dist/index.js"
# build the MCPB bundle
npm run build && npx @anthropic-ai/mcpb pack
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。