Local · MIT · v1.0

Your loyal Magpie
Lands every visual in the nest.

A gallery your AI agent curates for you, version after version.

Your LLM gives you chat history. Magpie gives you a local library for every non-code thing it makes — mockups, diagrams, charts, status reports, plans. Offline. Yours.

$npx magpie-mcp # opens localhost:3737 · auto-bumps port

Read the guide→

not a generator not a whiteboard not cloud not an LLM

What it is

An organizer, not a generator.

Magpie is a local dashboard plus an embedded MCP server. Your AI coding agent (Cursor, Cline, or any MCP host) writes the visuals; Magpie files them by project, tags them, versions them, lets you compare any two side-by-side. No LLM lives in the server. No cloud. No accounts.

It is

A library for the visuals your AI agent makes. Like Eagle or PureRef, but your AI fills the nest through MCP.

It isn't

A generator. The server has no model. Your AI agent does the drawing.

It isn't

A whiteboard. No infinite canvas, no drawing tools, no real-time collaboration.

Three actors

Master trains. Magpie tends.
Nest keeps.

You are the reader they all serve. Magpie is the bird; the Nest is your disk; the Master is the skill your AI gets.

trainer

Master

Teaches your AI Magpie's rules.

A portable Agent Skill bundled in the npm tarball — pure judgment, no code.

trained bird

Magpie

Tends the nest, version after version.

Your AI with the Master loaded. Brings visuals to the Nest, arranges by project, keeps versions in order.

dumb infra

Nest

Local. Quiet. Yours.

One SQLite file. One blobs folder. Lives at ~/.magpie/ on your disk. No LLM. No telemetry.

Master trains·Magpie tends·Nest keeps.

Format coverage

Seven formats. One nest.

All renderable, all comparable side-by-side, all offline.

HTML html

.html · sandboxed iframe

UI mockups, dashboards, landing pages — sandboxed iframe.

Mermaid mermaid

.mmd · bundled v11

Flowcharts, sequence, state, ER, gantt — bundled v11.

SVG svg

.svg · direct

Icons, logos, single vector elements — sanitized inline.

Markdown md

.md · marked + DOMPurify

Prose, READMEs, design notes — marked + DOMPurify.

Graphviz dot

.dot · graphviz-wasm

Directed and undirected graphs via WASM Graphviz.

Vega-Lite vega-lite

.vl.json · vega SSR

Bar, line, scatter, area charts — inline data only.

D2 d2

.d2 · @terrastruct/d2 wasm

Modern diagram-as-code with dagre or elk layouts.

—

+ more

deferred · permanently

PlantUML / Excalidraw break the offline + pure-JS rule. Why →

Three ways

Hand work back to your agent.

Same goal — get the bird flying with context loaded. Pick the channel that fits how you work.

Pillar 1 · Async

Outbox

Push a templated prompt to a server-managed queue. Your AI picks it up via magpie://inbox when it's ready. No timing dance.

Best for: batched work · overnight runs

Pillar 2 · Reusable

Templates

Design a prompt once with typed variables and visual references. 10 built-ins ship; scope your own to a project or visual.

Best for: recurring iteration loops

Pillar 3 · Direct

MCP Prompts

Same templates surface as slash commands in your AI host (/magpie:iterate). User-fired — no clipboard, no queue.

Best for: in-chat refinements

Quick start

Three minutes from `npx` to your first visual.

No accounts. No API keys. No build step.

Boot the server.
Run npx magpie-mcp · prints a dashboard URL · opens automatically.
Install the Master skill.
npx magpie-mcp --install-skill · drops the skill into Claude Code / Claude Desktop.
Wire your MCP host.
npx magpie-mcp --print-config · paste the snippet into your MCP host config, restart.
Ask for a mock.
"Mock a settings page" · the visual lands in the dashboard within a second.