Talorys deploys a personal AI agent inside your Cloudflare account

Running your own AI assistant has meant renting a server, picking a model host and leaving an account open with someone. Talorys reached 105 points on Hacker News on Friday morning Pacific time with a single command, npx create-talorys@latest, that deploys a personal agent inside the reader's own Cloudflare account. The repository was created the same day at 09:46 UTC and had 108 GitHub stars four hours later.

Talorys is open source under the MIT licence and runs on four Cloudflare services the installer lists as free-plan: Pages for the React frontend, Workers for a private API, Durable Objects for all data and scheduling, and Workers AI for the chat model. The project's README names the model it uses, @cf/zai-org/glm-4.7-flash, and says Workers AI is "yes, daily allocation" on the free plan.

What one install sets up

The installer checks for Node.js 20.18 or newer, opens Cloudflare's own OAuth page in the browser if the reader is not signed in, lets them pick an account and name the agent, and asks for an owner password that is hashed locally with PBKDF2-SHA256 and stored only as a Cloudflare secret. It then deploys a private Worker and a Pages project, binds them through a service binding so the Worker has no public URL, and prints the live pages.dev address. The README says the install verifies the deployment, including an unauthenticated-rejection check, without running any paid inference.

Interrupted installs resume on a second run, with the same resource names and no duplicate projects. Password reset is one further command: npx create-talorys@latest reset-password replaces the secret and signs out every device.

What runs when the daily AI allocation is used up

The README is specific about what Talorys does when the Workers AI quota resets. Chat shows a message and resumes the next day. Tasks, notes, memories and reminders keep working, because they do not call the model. Simple reminders and task digests never use AI at all, which keeps the daily reminder loop running on the free plan even on a heavy day.

The project ships adjustable per-request limits in the UI: a cap on output tokens, a cap on context tokens with older history summarised when it is reached, a cap on tool calls and reasoning steps per request, a daily cap on AI requests and a daily cap on scheduled AI runs. A usage panel shows local estimates and links to the Cloudflare dashboard for the real figures.

Single user, no telemetry, your Cloudflare

Talorys is designed for one person. There are no accounts, no teams and no signup. Sign-in is the owner password set at install time. The author states on the README that Talorys sends nothing to its developers: no telemetry, no analytics, no tracking, no advertising code. Data stays in a single SQLite-backed Durable Object in the reader's Cloudflare account, which Cloudflare's own privacy policy and Workers AI terms then govern.

The agent itself handles chat with streaming responses, durable memories a reader can view and edit, task and note CRUD from the UI and from chat, and automations such as one-time and recurring reminders. Scheduling runs on Durable Object alarms, so nothing has to stay online for the reminder to fire.

For a developer or a small team that wants a private, scoped assistant on data they own, Talorys turns the setup from a week of plumbing into one command, and the ceiling is the Cloudflare free plan rather than another company's trial.

Source

Talorys on GitHub. Hacker News discussion. Star count, licence, feature list, model name and install flow are all taken from the project's own README. Star and commit times are from the GitHub API.