@diagrammo/dgmo-mcp

Turn a conversation into a real diagram — without leaving your AI tool.

This MCP server gives Claude (and any MCP-compatible AI tool) the ability to render sequence diagrams, flowcharts, ER diagrams, C4 architecture, gantt charts, and 40+ other chart types from concise text markup — then hand the result off to a full editor for refinement. Ask for a diagram in chat; get a real one back.

What you can do

Ask in plain language — "diagram the auth flow as a sequence", "chart the Q3 plan as a gantt", "draw our services as a C4 diagram" — and Claude writes the markup and renders it. The markup stays readable and diffable:

flowchart Mutiny Resolution
direction-tb

[Sail]     Set sail under the captain
{Trouble?} Discontent in the crew?
{Vote}     Crew vote called
[Mutiny]   Seize the ship

(Sail) -> (Trouble?)
(Trouble?) -Yes-> (Vote)
(Vote) -Mutiny-> (Mutiny)

→ renders to the flowchart above. All rendering happens locally — no diagram data leaves your machine.

Tools

The three space tools send the diagram to Diagrammo Cloud. They need a space API key (minted on the Diagrammo account page) as DGMO_API_KEY in the server's environment — never as a tool argument, so the key stays out of the transcript.

share_diagram and open_in_app are the bridge out of chat: a diagram Claude generates becomes something you can refine, restyle, and embed — see below.

The four marked no open a browser or launch the desktop app. Over HTTP that would happen on the machine running the server rather than on yours, so they are not offered there — see Serving over HTTP.

Beyond the MCP server

The MCP server is one entry point into Diagrammo — a whole ecosystem built on the same DGMO markup. Generate in chat, refine in a real editor, embed anywhere:

• diagrammo.app — the desktop app. open_in_app drops an AI-generated diagram straight into it, with live preview, palettes, and export.
• online.diagrammo.app — a full editor in the browser, zero install. share_diagram URLs open right here.
• Docs integrations — drop DGMO fenced code blocks into your docs site: remark-dgmo, astro-dgmo, docusaurus-plugin-dgmo, fumadocs-dgmo.
• Obsidian — the Diagrammo Diagrams community plugin renders DGMO in your vault.
• CLI — npx @diagrammo/dgmo-cli file.dgmo -o out.png, or install it with npm install -g @diagrammo/dgmo-cli (macOS and Linux).

One markup, everywhere. A diagram you generate here renders identically in the app, in your docs, and in Obsidian — because they all speak DGMO.

→ Try it free at diagrammo.app

Setup

Easiest — one command

Install the dgmo CLI and let it wire everything up:

npm install -g @diagrammo/dgmo-cli   # macOS and Linux
dgmo install                         # auto-detects Claude Code, Codex, Claude Desktop, Cursor, …

Prefer Homebrew or pacman? See diagrammo.app/dev.

dgmo install configures each detected assistant non-interactively and points it at dgmo mcp, so there's no separate package to install or prompts to answer. Target one surface with dgmo install claude-code (or codex, claude-desktop, …).

Manual configuration

Prefer to edit configs yourself? Point any MCP client at the server via npx (no global install needed):

Claude Code — .claude/settings.local.json; Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "dgmo": {
      "command": "npx",
      "args": ["-y", "@diagrammo/dgmo-mcp"]
    }
  }
}

If you have the dgmo CLI installed, { "command": "dgmo", "args": ["mcp"] } works too. Restart the client after saving — the tools appear automatically.

Serving over HTTP

The setups above launch the server as a child process and talk to it over its standard input and output. That needs the server and the client on the same machine. Where they are not — a hosted agent platform, a container, one server shared by several people — start it as an HTTP endpoint instead. The same endpoint also suits one machine, when you want a single server always running for every AI tool you use:

dgmo mcp --http                              # with the dgmo CLI installed
npx -y @diagrammo/dgmo-mcp --http            # http://127.0.0.1:3333/mcp
MCP_TRANSPORT=http MCP_PORT=8080 npx -y @diagrammo/dgmo-mcp

Every option takes a flag or an environment variable, whichever your setup can express:

Each request is served independently — no sessions, nothing kept between calls — so one endpoint can serve several clients at once.

The server has no authentication of its own. It binds loopback by default and rejects requests carrying a Host header it was not told to expect, or an Origin header other than its own unless --allow-origin names it, which is enough for a client on the same machine or inside the same container. Anything reachable from a wider network needs your own authentication in front of it, and --allow-host for the hostname it will be reached by. Binding a non-loopback interface without naming a host prints a warning saying so.

--help prints all of this from the installed version.

Keep it running

Run the server as a background service so it starts at login and restarts if it stops. Change the dgmo path to what which dgmo prints.

Linux — systemd user service, saved as ~/.config/systemd/user/dgmo-mcp.service:

[Unit]
Description=Diagrammo MCP server over HTTP
After=network.target

[Service]
ExecStart=%h/.local/bin/dgmo mcp --http
Environment=PATH=%h/.local/bin:/usr/local/bin:/usr/bin:/bin
Restart=on-failure
RestartSec=3

[Install]
WantedBy=default.target

systemctl --user daemon-reload
systemctl --user enable --now dgmo-mcp

macOS — launchd agent, saved as ~/Library/LaunchAgents/app.diagrammo.mcp.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key><string>app.diagrammo.mcp</string>
  <key>ProgramArguments</key>
  <array>
    <string>/opt/homebrew/bin/dgmo</string><string>mcp</string><string>--http</string>
  </array>
  <key>EnvironmentVariables</key>
  <dict><key>PATH</key><string>/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin</string></dict>
  <key>RunAtLoad</key><true/>
  <key>KeepAlive</key><true/>
</dict>
</plist>

launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/app.diagrammo.mcp.plist

Then point each AI tool at http://127.0.0.1:3333/mcp as a streamable-HTTP MCP server.

Privacy

All rendering is local. Your diagram markup and the images it produces never leave your machine, except when you explicitly call share_diagram (which encodes the diagram into a diagrammo.app URL). See the privacy terms.

Dev hub (AI-tuning tools)

pnpm hub

One command, one server, one browser tab. The hub opens a tabbed shell over the three AI-tuning dev tools — switch between them with the top tabs, no separate ports or commands to remember:

• Trigger tuning — edit the phrase/concept vocabulary that drives suggest_chart_type, score prompts live, save back to triggers.json.
• LLM judge — judge chart-type descriptions against prompts with claude -p.
• Guidance studio — author the per-type styling guidance the server delivers (the <!-- TIPS --> blocks in dgmo's language-reference.md, sliced into get_language_reference): pick a type, edit how the AI is told to style it, run a prompt against a committed dataset fixture (so inputs never move between runs), and see the generated DGMO + rendered image side by side. The picker doubles as a coverage bar; "Compare 3×" renders no-guidance vs your tips for a by-eye check; Save validates and writes back to language-reference.md.

These tools are dev-only and never bundled into the published server. (The standalone pnpm harness and pnpm studio scripts still run a single tool each if you ever want one in isolation.)

Contributing & releases

Development setup and the release workflow live in CONTRIBUTING.md.

License

MIT