Quixi

Your AI history, independent of the company that generated it.

Quixi is being rebuilt around a provider-neutral local archive shared by desktop and web. The product specification describes the destination; the architecture guide defines repository boundaries.

Current status

The shared React app now has a durable conversation library, ChatGPT/Claude import, lexical search, provider setup, streaming chat, and edit/regenerate branches. Controlled HTTP tests exercise both provider adapters and browser-process restart through the production SQLite WASM/OPFS worker. Archive export/restore and the embedding scheduler are being integrated; the full roadmap is unfinished. See application evidence, the implementation roadmap, and the platform matrix for remaining gates.

The original runnable Gemma/Metal application is preserved intact in legacy/prototype, with its own Cargo workspace, lockfile, kernels, vendor patch, build script, and documentation.

Develop

Use Node.js 22.13 or newer and npm. Building the pinned SQLite distribution also requires Python 3.12+, curl, and Docker with Linux containers. Desktop development requires Rust 1.92+ and the platform's Tauri prerequisites.

npm ci
npm run sqlite:build
npm run dev

The web application is available at http://127.0.0.1:5173. Visit /storage-proof for the developer harness. SQLite binaries are generated, ignored artifacts; build them once before frontend development or checks, and rebuild when the pinned sources change. npm run sqlite:verify checks their hashes and executes the SQL smoke tests. CI builds this same artifact on Linux and verifies it again in each consuming job. Docker builds it in a pinned SDK stage without requiring generated files in the build context.

npm run dev:desktop
npm run check
npm run check:desktop
npm run build:desktop

check typechecks all TypeScript source and builds both frontend entry points. check:desktop also checks the native host. build:desktop builds and bundles the desktop application for the current platform; signing is not configured yet. Cross-platform storage support remains subject to A1.

Desktop provider setup uses the native registered OpenAI and Anthropic endpoints. The browser requires an operator-configured relay; without one, local import, browsing and text search remain available. Set VITE_QUIXI_RELAY_CONFIG at build or development-server startup to a JSON object containing origin, operator, privacy (self_hosted_remote, custom_remote or quixi_relay) and destinations: {"openai": "registered-server-id", "anthropic": "registered-server-id"}. Use an HTTPS origin. This configuration is public routing metadata; enter provider and relay credentials through the Providers screen, never in build variables. The relay service must independently register the matching destinations. Optional regional entries add US/Europe connections with a server-computed configuration identity; each can name a separate reviewed relay origin/operator. See the regional relay setup for provisioning and the operator trust boundary.

Browser integration checks run against production bundles:

npx playwright install chromium webkit
npm run test:app:browser
npm run test:e2e

Playwright WebKit results do not establish installed Safari or Tauri WebView support. Browser reports and traces are generated under test-results/.

The independent encoder has a separate provisioning and validation entry:

python3 packages/quixi-embed/native/ci.py --browsers

It requires uv, clang, Docker, and Node, and provisions its pinned Python/model inputs. Add --full for all scalar numerical and retrieval gates. CI runs the smoke/browser checks on pushes and pull requests; a manual workflow run includes the full scalar gate. See QuixiEmbed for the runtime's current coverage and optimization work still outstanding.

Layout

Path	Responsibility
apps/web	Browser entry point and browser host adapter
apps/desktop	Desktop entry point and Tauri privileges
apps/extension	Future provider extraction and bundle transfer
packages/app	Shared UI and application workflows
packages/core	Canonical model, contracts, and pure rules
packages/storage	Storage Worker, SQLite WASM, OPFS, migrations, archives
packages/providers	Live provider adapters
packages/importers	Historical normalization, identity, provenance
packages/search	Shared chunking, retrieval orchestration, rank fusion
packages/documents	Document worker and text normalization
packages/quixi-embed	Independent embedding runtime and model compiler
services	Relay and planned optional Cloud services
deploy	Deployment configuration
tests	Cross-package integration, host tests, shared fixtures
perf	Product-scale benchmarks
tooling	Build and release orchestration
legacy/prototype	Original Gemma application, outside the active workspaces

Run the original prototype

On Apple silicon with the Metal toolchain:

cd legacy/prototype
cargo test --workspace --locked
cargo run --locked

The prototype still downloads its pinned model on first launch. Its history and inference behavior are unchanged. It is not the implementation of the new spec.