Luia MCP Server
Enables Claude to access and update a team's design knowledge stored in a local Markdown folder, including searching decisions and recording new conventions.
README
<img src="docs/preview.jpg" alt="The Luia graph: team knowledge at the centre, project work fanning out around it, with one node selected to show its connections" width="100%">
Luia
A knowledge graph for design teams.
Design teams accumulate knowledge faster than they can organise it. A spacing rule agreed in March, a carousel pattern that tested badly, the reason a component was built one way and not another. It ends up scattered across docs, threads, and people's memory, and by the time someone needs it nobody can find it.
Luia has two halves:
- An MCP server that plugs into Claude. Designers and developers install it once, and from then on Claude reads the team's conventions before answering and writes new decisions back as they are made.
- A graph viewer that renders the same knowledge spatially, so you can see how decisions connect instead of scrolling a folder tree.
Both read the same plain Markdown folder. There is no database, no backend, and no service to sign up for — which means a team shares its knowledge by putting that folder in git.
Install the MCP server
Requires Node.js 20.19+ or 22.12+.
git clone https://github.com/gustavocambareri/Luia_MCP.git
cd Luia_MCP
npm install
Point Claude at it. LUIA_KNOWLEDGE_DIR is the folder your team's Markdown
lives in — anywhere you like:
claude mcp add luia --scope user \
--env LUIA_KNOWLEDGE_DIR="$HOME/luia-knowledge" \
-- node "$PWD/server/index.mjs"
Verify it connected:
claude mcp list # luia: … - ✔ Connected
For Claude Desktop, add the same thing to claude_desktop_config.json:
{
"mcpServers": {
"luia": {
"command": "node",
"args": ["/absolute/path/to/luia/server/index.mjs"],
"env": { "LUIA_KNOWLEDGE_DIR": "/absolute/path/to/your-knowledge" }
}
}
}
That's the whole setup. Ask Claude "what's our spacing rule?" and it searches the store before answering.
What Claude can do with it
| Tool | When Claude reaches for it |
|---|---|
search_knowledge |
Before answering anything about how your team designs or builds — so the answer matches decisions you already made |
get_document |
When a search excerpt looks relevant and it needs the full rule |
record_decision |
The moment you settle a convention, so it survives the session |
list_documents |
To see everything the team has written down |
The two that matter are search and record. Search is described so Claude calls it unprompted at the start of design work — nobody remembers to ask for their own conventions. Record is a single call with no ceremony, because a capture step with friction is a capture step that doesn't happen.
Using it with your team
The knowledge store is a folder of Markdown files, so sharing it is just git:
cd ~/luia-knowledge
git init && git add . && git commit -m "Team knowledge"
git remote add origin git@github.com:your-team/design-knowledge.git
git push -u origin main
Teammates clone that repo and point their own LUIA_KNOWLEDGE_DIR at it. Pull
to get everyone's decisions; push to share yours. Decisions arrive as readable
diffs you can review like any other change.
Each document is plain Markdown with frontmatter — editable by hand, no tool required:
---
title: Spacing System
project: atlas # optional; omit for team-wide knowledge
author: your-name # or 'team' for collectively-owned decisions
tags: [spacing, layout]
created: 2026-03-09
---
8px base unit, with a 4px step for tight UI.
**Why:** the ratio between gaps carries more meaning than absolute values.
[!TIP] Always write the why. A decision without its rationale is one the team relitigates in six weeks — and the rationale is what makes Claude apply the rule correctly to a case you didn't anticipate.
The graph viewer
- Spatial navigation — a 3D force-directed graph you can pan, zoom, and explore, with hand-composed positions so the layout reads deliberately rather than like a hairball.
- Four ways to filter — by scope, project, author, or free-text search across titles, tags, descriptions, and section headings.
- The author lens — dim everything except one person's contributions to see who holds which knowledge, and where a single point of failure is forming.
- Typed relationships —
foundation,reference, andsiblingedges are drawn differently, so how two documents relate is legible at a glance. - Readable content — every node carries full Markdown, a description, and a section outline; click any node to read it without leaving the graph.
Running the viewer
npm run dev
Open the address Vite prints (usually http://localhost:5173) and the demo
graph loads immediately.
npm run build # production build to dist/
npm run preview # serve that build locally
npm run lint # eslint
[!NOTE] The bundled graph is entirely fictional. Meridian, Lumen, Atlas, Cadence, Harbor, and Verso are invented engagements, and every person credited in them is made up. Nothing here is real client work.
Viewer data format
The viewer reads src/data/graph-data.json. Replace it with your own and the
app is yours.
A node looks like this:
{
"id": "projects/atlas/block-library",
"name": "Atlas — Block Library",
"slug": "block-library",
"scope": "project",
"project": "atlas",
"author": "your-name",
"tags": ["atlas", "blocks", "templates"],
"created": "Mon Mar 16 2026 01:00:00 GM",
"fileSize": 8200,
"description": "Reusable blocks and page templates for the Atlas redesign.",
"body": "# Atlas — Block Library\n\n## Navigation\n\nSticky header that…",
"sections": ["Atlas — Block Library", "Navigation"],
"connections": ["_team/design-system-patterns"]
}
And an edge connects two of them:
{
"source": "_team/design-system-patterns",
"target": "projects/atlas/block-library",
"type": "reference",
"label": "informs",
"description": "Team patterns applied in the Atlas block library."
}
The full shape is defined in src/lib/types.ts.
Fields worth understanding
| Field | What it does |
|---|---|
scope |
team for shared knowledge, project for engagement work. Drives colour and grouping. |
project |
Required when scope is project. Becomes a filter chip automatically. |
author |
Powers the author lens. Use team for collectively-owned documents. |
fileSize |
Controls node radius. Scale your values into roughly 2600–19500 so sizes stay differentiated. |
body |
Markdown, rendered in the detail panel. |
sections |
Heading list, included in search. |
type |
Edge weight and curvature: foundation arcs through the centre, sibling hugs the perimeter. |
Adding a project
- Add your nodes to
graph-data.jsonwith a newprojectvalue. - Add a colour for it in
src/lib/colors.ts→PROJECT_COLORS. Without one it falls back to the default ink, and becomes indistinguishable from team nodes. - Optionally place its nodes in
MANUAL_POSITIONSinsrc/components/graph3d/useForceLayout.ts. Nodes without coordinates are positioned automatically, but hand-placing them keeps clusters legible.
The filter chips read the project list straight from your data, so step 1 is enough to make it appear in the UI.
[!TIP] Keep
idvalues path-like (projects/<project>/<slug>). Nothing enforces it, but it keeps the file browsable and makesconnectionseasy to write by hand.
Design decisions
A few choices are deliberate and worth knowing before you extend it:
- Positions are authored, not simulated. A pure force layout drifts on every
reload and buries the structure. Coordinates live in
useForceLayout.tsso the composition is stable and intentional. - Colour is a system, not decoration. Every hue in
colors.tsis derived from a five-colour palette and tuned to a narrow contrast band, so no project reads as visually heavier than another. - Uppercase titles, explicit weights. Type rules are applied consistently
across the canvas and panel; if you add UI, set
fontWeightexplicitly rather than inheriting it.
Project structure
server/ # the MCP server
├── index.mjs # tool definitions + stdio transport
└── knowledge.mjs # Markdown store: read, write, search
src/ # the graph viewer
├── data/graph-data.json # demo content — replace this with yours
├── lib/
│ ├── types.ts # GraphNode, GraphEdge, GraphData
│ └── colors.ts # palette, scope/project/author colour maps
├── hooks/useGraphData.ts # filtering, search, selection state
└── components/
├── FilterBar.tsx # scope/project/author filters, search
├── DecisionLog.tsx # per-node detail panel
└── graph3d/ # the 3D scene
├── useForceLayout.ts # hand-placed node coordinates
├── GraphNode3D.tsx # node + label rendering
└── GraphEdge3D.tsx # typed, curved edges
Built with
React 19 · TypeScript · Vite · Three.js via React Three Fiber · Tailwind CSS
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
E2B
Using MCP to run code via e2b.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.