Talorys runs a personal AI agent entirely on Cloudflare's free tier

An open-source project called Talorys, which reached the Hacker News front page via its GitHub repository, offers a personal AI assistant that runs entirely inside a Cloudflare account you control. According to the project's README, the complete stack — web frontend, API, storage, scheduling and model inference — is built from services available on Cloudflare's free plan, and deployment is triggered by a single command: npx create-talorys@latest.

What the agent does

Talorys is deliberately single-user: there is no signup, no team functionality and no service operated by the developers. The installer sets one owner password, and any number of devices can sign in as that same owner.

Feature-wise, it provides streaming chat with Markdown rendering and indicators showing when the model is calling tools; a durable memory of personal facts and preferences that the owner can view, edit and delete, with only the most relevant memories injected into each prompt; and full task, note and project management, available both through the UI and through conversation. It also runs automations — one-time and recurring reminders, daily task digests and optional AI routines — delivered to an in-app notification center. These are driven by Durable Object alarms, so nothing has to stay online between firings.

A notable design decision is that the non-AI parts keep working when the AI does not. If Workers AI is unavailable or the free daily allocation is spent, tasks, notes, memories and reminders continue to function; only chat pauses until the reset.

How it works

The architecture keeps the exposed surface small. The browser only ever talks to the deployed pages.dev site; /api/* requests are handled by a Pages Function that forwards them over a service binding to a private Worker running a Hono router. That Worker is deployed with its workers.dev subdomain and preview URLs disabled, so it has no public address of its own, and authentication happens in the Worker rather than the frontend. Chat responses stream end-to-end as Server-Sent Events.

The agent itself is a Durable Object built on Cloudflare's Agents SDK, with SQLite inside it storing conversations, memories, tasks, notes, projects, automations, sessions, settings and usage. The model is Workers AI's @cf/zai-org/glm-4.7-flash, used with streaming and tool calling. The project provisions no R2, D1, KV, Vectorize, AI Search or Workflows resources — just Pages, Workers, Durable Objects and Workers AI, all of which have free-plan tiers.

Deployment, updates and data

The installer checks for Node.js 20.18 or newer, uses a bundled Wrangler CLI, and authenticates through Wrangler's standard OAuth flow rather than a Global API Key. It asks for an owner password, hashes it locally with PBKDF2-SHA256 and stores only the hash as a Cloudflare secret, generates a 256-bit session secret, deploys the Worker and Pages project with the service binding, then verifies the live deployment without running any inference. An interrupted install can simply be re-run: it reconciles existing resources, never duplicates them and never deletes data. Non-interactive installs are supported through environment variables.

Updates go through the same command with an update argument; migrations are append-only, the Durable Object namespace is never recreated, and the owner password and sessions survive upgrades. A settings panel can export a JSON backup of conversations, memories, tasks, notes, projects, settings and automations, and re-importing the same file twice is harmless. Local development uses a deterministic mock AI provider, so no Cloudflare login is needed to hack on the code.

Costs, quotas and privacy

Talorys is engineered to fit the Workers Free plan and never enables paid features on its own, but the README is explicit that this is not unlimited free computing. Free plans carry account-level quotas for requests, Durable Object usage and a daily Workers AI Neuron allocation, all set by Cloudflare and subject to change. When the AI allocation runs out, chat displays a message and resumes after the daily reset. Adjustable guardrails — caps on output and context tokens, tool calls, reasoning steps, daily AI requests and scheduled AI runs — help keep usage in check, and plain reminders and task digests never invoke the model at all.

On privacy, all data lives in one SQLite-backed Durable Object in your own account, and the project ships no telemetry, analytics or tracking code. The README still acknowledges that Cloudflare processes your data to provide the service, including running inference on your chat messages and the memories attached to each prompt, and directs readers to Cloudflare's privacy policy and Workers AI terms.

Why it matters

Talorys is a working demonstration that a reasonably full-featured personal agent — chat, memory, task management and scheduled automations — can be assembled entirely from an edge platform's free tier, with no server to babysit and no vendor-run backend behind it. For users wary of commercial assistants, it shifts the trust boundary: your data sits in infrastructure you control, though inference still runs on Cloudflare's model serving. The single-user scope is a deliberate limitation rather than an oversight, and the frank documentation of quotas, including graceful degradation when the AI allocation is exhausted, makes the free-tier promise concrete instead of aspirational.