Linkedin Network Constellation

See who you know. Drop in your LinkedIn connections and they arrange themselves by what people do: thousands of names become a few dozen clusters you can orbit, search and question in plain English. It runs entirely in your browser. Nothing is uploaded.

Try it → Start with the built-in demo of 250 invented people, or drop your own export in.

LinkedIn search answers "who matches this keyword". It doesn't answer "who do I already know in PropTech?", and that is the question that matters when you need an introduction.

• See the shape of your network. Every connection is placed by the field their job puts them in, and joined to any employer they share with someone else. Each field is its own cluster in its own colour.
• Ask it a question. Anyone in satellite imagery licensing? is answered on your machine in milliseconds, with no model and no API key. The query it ran is shown above the answer, so you can check its reasoning.
• Open the full picture. Click anyone for their headline, employer and everyone else you know there. Click an employer for who works there.
• Go deeper, if you want. With your own Anthropic API key, Claude looks up where employers are based, so any VC based in London? can be answered. It can also research one person on the public web when you press Enrich profile.

Get started

1. Export your connections. On LinkedIn, open Get a copy of your data, tick Connections and request the archive. It arrives by email, usually within ten minutes. This is your own data, so there's no scraping and no session cookie.
2. Drop Connections.csv on the page. It is read and classified in that tab, and kept in that browser.
3. Ask it something, or just orbit it.

Your first session →	A ten-minute walkthrough of everything, on the demo, with screenshots.
Install and self-host →	Run it locally, build a single file, put your own copy on GitHub Pages, and troubleshooting.
How questions are answered →	How a question becomes a filter, with measured results.
Ask from your AI tools (MCP) →	Use your network from Claude, Cursor, VS Code and other MCP clients.

Ask from your AI tools (MCP)

The same engine also runs as an MCP server, so Claude, Cursor, VS Code and other AI tools can answer questions about your network in the chat, like "who do I know at Esri?" or "which three VCs do I know best?".

Claude Desktop: download network-constellation.mcpb, double-click it, choose your Connections.csv (not the .mcpb itself), and start a new chat. Everything else, for example Claude Code:

claude mcp add network-constellation --scope user -- npx -y github:asturksever/network-constellation --csv /full/path/to/Connections.csv

It runs on your machine and reads only that file, but what it returns (the names, titles and employers of the people a question matches) is sent to your AI tool's provider as part of the conversation. It never returns email addresses, and returns profile links only if you ask it to. Setup for every client, and the six tools: docs/mcp.md.

Privacy

Your CSV is read in the tab you dropped it into and kept in that browser's storage (IndexedDB). It is never uploaded. There is no backend, no account and no analytics. Forget in the control panel erases everything the page kept.

Without an API key, the page makes one request to anyone else: it loads the 3D library from cdn.jsdelivr.net.

With your own Anthropic key, three things can be sent to Anthropic, and nothing else:

1. Employer names, to look up headquarters and organisation type. Sent when you press Look up employers in Settings, or Label on one employer.
2. The text of your questions, each one you ask while a key is set, so Claude can read it more closely. Anything it adds is marked in the answer.
3. One person's name, headline and employer, only when you press Enrich profile on their panel, so Claude can search the public web for them. This is the only thing that ever sends a name, and it never runs by itself.

Never email addresses or profile links. Requests go straight from your browser to api.anthropic.com, and the page's Content-Security-Policy lets its scripts send data nowhere else. If a brief finds a public photo, your browser loads it from the site that hosts it, as it would any image. Results are kept in your browser, so nothing is paid for twice. Your key is kept only if you tick Remember. Use a key with a spend limit.

What that costs: about $0.06 to place the employers that two or more of your connections share, in a 10,000-person network. About $0.70 to place all of them. $0.10–$0.30 for each Enrich profile.

The MCP server is separate from the page and has its own trade-off, described above and in docs/mcp.md.

npm run logos is a separate, optional build step on your own machine. It sends domain guesses made from employer names to unavatar.io, DuckDuckGo and Google to fetch favicons. Company names only, never people's.

What this is honest about

• The lines are memberships, not relationships. A line joins a person to a field or to an employer, never to another person. LinkedIn doesn't share who knows whom, so this is a map of what people have in common, not a social graph.
• Location means the employer's headquarters, not where the person lives, and it exists only after enrichment. A location question leaves out anyone whose employer could not be placed, and says how many that was.
• Roles are self-descriptions, not verified titles.
• Employer is what people state. The official export carries one for most people; built from headlines alone it was about a third. A blank means they didn't state one.
• Roughly a third state no seniority. "Unstated" is a real category, not missing data to be guessed at.
• The classifier is patterns, not a model. See src/taxonomy.js: it is readable and editable on purpose. The same rules that classify a job title also parse a question.
• An enriched profile can be about the wrong person. Names collide. The brief says how confident it is and cites its sources. The prompt treats a confident brief about a namesake as the worst possible outcome.

Run it yourself

Node 18 or newer. There are no dependencies, so there is nothing to install.

git clone https://github.com/asturksever/network-constellation.git
cd network-constellation
npm run dev            # http://localhost:8080

npm test               # the tests
npm run pages          # the site as GitHub Pages serves it, on :8090
npm run bundle:demo    # one self-contained HTML file with the demo, safe to share
npm run mcpb           # the Claude Desktop extension, dist/network-constellation.mcpb
npm run bundle         # the same with YOUR graph inside: keep it to yourself

The install guide covers deploying your own copy, other hosts, and what to do when something goes wrong.

How it is built

A static page with no framework, no bundler and no dependencies. The ES modules run the same in the browser and in Node, which is how the tests run.

index.html          the page, the landing and the Content-Security-Policy
src/taxonomy.js     the classification rules: edit here when a field is wrong
src/classify.js     headline -> role, company, seniority, field
src/csv.js          CSV reading, including Excel's encoding and wrong-file detection
src/build.js        rows -> the graph (shared by Node and the browser)
src/ask.js          question -> filter -> ranked shortlist, no model involved
src/graph.js        the 3D scene: nodes, links, colour, camera
src/palette.js      generated colours for every field
src/labels.js       field labels projected from 3D
src/highlight.js    the marker that locks onto a found person
src/logos.js        employer logos, when you have fetched them
src/dom.js          the small DOM helpers every UI module shares
src/ui.js           the control panel, tooltip and status line
src/overview.js     the network overview at the top of the panel
src/askui.js        the question box and its answer panel
src/detail.js       the person and employer panels
src/upload.js       the landing page, the drop and the column mapper
src/store.js        IndexedDB: your graph, employer facts and research briefs
src/llm.js          the optional Anthropic calls, with your key
src/enrich.js       employer names -> headquarters and organisation type
src/askllm.js       question -> extra filter constraints
src/research.js     one person -> a sourced brief, on a button only
src/enrichui.js     the Settings sheet
src/main.js         boot: find a graph or show the landing, then wire it all
mcp/server.mjs      the MCP server: JSON-RPC over stdio, six read-only tools
mcp/network.mjs     what those tools do, over the same build and ask modules
mcp/icon.png        the Claude Desktop extension's icon
scripts/            the dev server, the data build, logos and the bundler
test/               node:test, no framework

Built on 3d-force-graph.

Contributing

See CONTRIBUTING.md. The most useful change is to src/taxonomy.js: it is wrong in ways only someone with a different network can see. If your field comes out as "Other", that is the file to fix.

Licence

MIT. See LICENSE.

Network Constellation is not affiliated with or endorsed by LinkedIn.