LocalLink: let websites use the visitor's local AI models

LocalLink lets a website use AI models running on the visitor’s own computer. The visitor picks “Local model” on the site and approves it once. The site can then list and chat with their local models: Ollama, LM Studio, llama.cpp, Jan or any OpenAI-compatible server.

Draft v0.1. Tested against a mock OpenAI-compatible server; not yet verified against real Ollama or LM Studio installs or from public HTTPS sites.

Components

Path	What it is
`SPEC.md`	The protocol, v0.1
`bridge/locallink-bridge.mjs`	The program the visitor runs. Zero dependencies, Node 18+
`sdk/locallink.js`	What a website imports
`demo/`	A sample chat site
`test/`	End-to-end test and a mock model server

Try it

Run each command in its own terminal, from a clone of kalaris-locallink:

node bridge/locallink-bridge.mjs
node demo/serve.mjs

Open http://localhost:8787, choose Local model on this computer and click Allow in the approval tab. With no model server installed, run node test/mock-backend.mjs and start the bridge with --backend mock=http://127.0.0.1:9911.

Use it on a website

import { LocalLink } from './locallink.js';

const link = await LocalLink.connect({
  appName: 'My App',
  onCode: code => showCode(code),   // the visitor matches this against the approval screen
});
const models = await link.models();  // [{ id: 'ollama/llama3.2:3b', backend, name }]
for await (const text of link.chat({ model: models[0].id, messages })) append(text);

LocalLink.resume() reconnects a returning visitor without a prompt.

Manage access

node bridge/locallink-bridge.mjs grants
node bridge/locallink-bridge.mjs revoke https://example.com

LocalLink is MIT licensed. For on-device inference research, see Anudhi.