Talorys is an open-source personal AI agent that runs entirely inside a single user's Cloudflare account. One command — npx create-talorys@latest — deploys a React frontend on Cloudflare Pages, a private Hono-based Worker with no public URL, and a SQLite-backed Durable Object that stores everything: conversations, memories, tasks, notes, projects, automations, and settings. The architecture is single-user by design. No signup, no teams, no accounts managed by the developers. The owner sets a password during installation, hashed locally with PBKDF2-SHA256 and stored as a Cloudflare secret. The technical architecture is clean and deliberately constrained. The browser talks only to a .pages.dev site. API requests route through a Pages Function over a service binding to the agent Worker, which has no public URL (workersdev: false, previewurls: false). Authentication happens in the Worker, not the frontend. Chat uses Workers AI's @cf/zai-org/glm-4.7-flash model with streaming responses via Server-Sent Events. Scheduled reminders and recurring routines run on Durable Object alarms — nothing needs to stay online. The free-tier discipline is notable. Talorys uses only Cloudflare Pages, Workers, Durable Objects with SQLite, and Workers AI — no R2, D1, KV, Vectorize, AI Search, Workflows, or any paid service. When the daily Workers AI neuron allocation runs out, chat shows a message and resumes at reset. Tasks, notes, memories, and reminders keep working without AI. Built-in guardrails cap output tokens, context tokens, tool calls, reasoning steps, and daily AI requests. The privacy model is the real product differentiation. All data lives in a single SQLite-backed Durable Object in the user's own Cloudflare account. No telemetry, analytics, tracking, or advertising code. Nothing is sent to the Talorys developers. The project explicitly notes that Cloudflare itself processes your data to provide its services, including running inference on chat messages — an honest disclosure that most self-hosted projects skip. The installer is idempotent and resilient. Interrupted runs can be resumed; duplicate resources are never created; passwords aren't re-prompted if already configured; data is never deleted. Updates redeploy the latest Worker and frontend to existing resources, with schema migrations running automatically and transactionally. Backup and restore work through the UI, producing a JSON file that can be imported without duplication. The project is MIT-licensed with three named contributors. Local development requires only Node.js 22+ and uses a deterministic mock AI provider, so no Cloudflare login is needed. Playwright UI tests and optional real-account deployment tests exist. The codebase is a monorepo: apps/agent (Worker), apps/web (React + Vite), packages/shared (schemas, types, recurrence logic), and packages/create-talorys (the installer). This is a genuinely generative open-source project. The value proposition is architectural: your data, your account, your agent. The developers extract nothing — no usage data, no analytics, no account system that creates lock-in. The constraint is Cloudflare's free tier itself, which is both the enabler and the ceiling.
