Architecture

How the pieces fit together — Zero sync, apps, flows, and AI

start.chat is built on a few core systems that compose together. Understanding how they connect helps when building apps and writing flows. The source and self-hosting release are not public yet; see Software Availability and Rights for the current status.

Stack

Layer	Technology	Role
Frontend	React + Tamagui	Universal UI (web, desktop, mobile)
Routing	One (Vite)	File-based routing, SSG/SSR/SPA
Sync	Orez	Real-time local-first data sync
Database	SQLite	Namespaced primary storage
Desktop	Tauri	Native macOS/Windows/Linux app
Mobile	React Native (Expo)	iOS and Android
AI	Multi-provider	Claude, Gemini, Qwen, Ollama, more
Flows	Pipedown	DSL for multi-step automation

Data flow

All data flows through Orez’s local-first sync runtime. The client maintains a local replica of the data it has access to. Mutations are optimistic — the UI updates immediately, then the server confirms or rebases.

client → optimistic mutation → local state updates instantly
       → server mutation → namespaced SQLite write → sync to all clients

This means:

The UI is never blocked waiting for the server
Offline edits queue and sync when connectivity returns
Conflict resolution is automatic
Every query is a local read — fast regardless of server latency

Permissions at the sync layer

Zero enforces access control at the data level through where predicates. When a client subscribes to data, the server filters rows based on the user’s roles and permissions. Private channels, secret data, and admin-only resources are invisible to unauthorized clients — they never leave the server.

Apps

Apps are the primary extension mechanism. An app definition (createApp()) declares:

Actions — functions that agents and flows can call
Backends — scheduled code that runs on a timer (runsEvery)
Blocks — React components that render inline in chat
Data — structured key-value storage scoped to the app
OAuth — connect to external services
Webhooks — receive events from external services

Apps compile in a sandbox. The sandbox injects a global namespace with tools — tools.data, tools.message, tools.generate, tools.strategy, tools.notify, and more. Apps cannot access the filesystem or make arbitrary network requests.

Flows

Flows are written in Pipedown, a plain-language DSL that compiles to an AST. The runtime executes steps sequentially or in parallel, calling app actions, LLM providers, or built-in operations at each step.

event trigger → parse Pipedown → execute steps → store results
                                    ↓
                              app actions, AI calls,
                              data writes, notifications

Flows can be triggered by events (new message, webhook, data change), schedules (every 6 hours), or manual commands. Each step has access to previous step outputs via tags.

AI

The AI system composes abilities into a system prompt at request time. When a user messages HUD:

Load channel config (model tier, system prompt, enabled abilities)
Assemble context (recent messages, available apps, user role)
Build system prompt from enabled ability definitions
Stream response, executing tool calls as they arrive
Post-process: citations, code blocks, action results

Model tiers

Seven tiers from xxs to xl, each mapping to specific models:

Tier	Use case	Examples
xl	Complex reasoning	Claude Opus, Claude Sonnet
lg	Detailed responses	Claude Sonnet, Claude Haiku
md	General conversation	Claude Haiku (extended thinking), Gemini Flash
sm	Tool-calling/strategy	Gemini Flash (high thinking), Claude Haiku
xs	Simple tasks	Gemini Flash, Claude Haiku
xxs	Quick operations	Gemini Flash Lite, Claude Haiku

Server admins set the default tier. HUD selects models within that tier based on task complexity. Ollama is supported for fully self-hosted AI.

Search

Search combines three ranking signals over namespace-local search documents:

Signal	How it works	Best for
Text	Token and phrase matching with title boost	Exact phrases and known terms
Recency	Time and document-kind weighting	Recent conversations and trusted documents
Semantic	Cosine similarity over stored embeddings	Conceptual queries, “find messages about…”

Results are merged with weighted ranking. Product mutations update search documents explicitly, so new content is findable without a database trigger.

Routing

File-based routing via One. Routes in app/ map to URL paths:

app/(chat)/ — authenticated chat UI
app/(site)/ — public marketing site, docs, blog (SSG)
app/api/ — REST API endpoints
app/auth/ — authentication flows

The (chat) group wraps everything in ~15 context providers (current server, channel, user, theme, Zero client, etc). The (site) group uses static generation for docs and blog pages.

Desktop sync

The Tauri desktop app adds bidirectional file sync. Your server’s apps and flows mirror to ~/start/{server-slug}/ on your filesystem. Edit in your editor, changes push back as drafts. Cloud edits write to local files.

Safety guards protect against accidents: mass change detection, large change ratio checks, freshness verification, offline protection, and per-file confirmation for deletions.

Features

Comparison