Koinos AI Core (kaiapp)

The desktop application for Koinos AI — local-first AI (chat + self-hosted OpenAI-compatible API on your own hardware) with optional provider earning on the decentralized compute network, settled in KAI on Koinos.

Status: M1 (local core) in progress. The Core service skeleton runs and is tested; the desktop UI, installer, and real-model verification are next.

core/                 Koinos AI Core service (zero runtime dependencies, Node ≥22)
  server.js           entrypoint: wires hardware detect, models, runtime, gateway
  lib/gateway.js      local OpenAI-compatible API (127.0.0.1): /v1/models,
                      /v1/chat/completions (streaming), /core/* control plane
  lib/runtime-manager.js + lib/runtimes/llamacpp.js
                      supervises llama-server as a child process (swappable runtime)
  lib/model-manager.js  hash-verified (sha256-pinned), resumable model downloads
  lib/keys.js         scoped API keys — hashed at rest, never wallet keys
  models/catalog.json versioned model packages + capability aliases
  scripts/pin-model.js  pins a package's sha256 into the catalog (run once, commit)
  test/               node --test suite incl. an integration chain against a fake
                      llama-server child process (16 tests)

ui/                   desktop UI — plain web app served by the gateway itself;
                      runs identically in the Electron shell or a browser tab
electron/main.js      thin sandboxed shell: boots Core in-process, opens a window
                      onto the gateway (window state persisted, external links
                      to system browser)

Run it:

npm test          # core tests need no dependencies; the browser test uses
                  # playwright-core + Chromium and skips itself when absent
npm run core      # headless: Core + UI on http://127.0.0.1:41100
npm start         # desktop shell (requires `npm install` for electron)

First run on a networked machine: node core/scripts/pin-model.js smollm2-135m-instruct-q8_0@1 --write to pin the dev model's hash (downloads refuse to run unverified — the catalog hash is the package's identity), and put a llama-server build at ~/.koinos-ai/runtimes/llamacpp/ (or point KAI_LLAMA_BIN at one).

• docs/V1_PLAN.md — the accepted V1 plan: milestones mapped to the spec's non-negotiables, architecture, and the module reuse map from Koinos-Node.
• Canonical requirements live in Koinos AI — Master Source of Truth (kept outside this repository); the plan cites it by section (§).

Related repositories: therexdev/kai (public website), therexdev/koinos-node (Koinos node manager / wallet / funding — the reuse foundation).