Talorys Runs a Personal AI Agent on Cloudflare's Free Plan

Talorys, an MIT-licensed personal AI assistant that deploys into a user's own Cloudflare account, drew 313 points on Hacker News on Oct. 10 for a simple promise: chat, notes, tasks and reminders with no server to rent. The catch is a daily Workers AI allowance of 10,000 Neurons, which Cloudflare's own price list converts into a few hundred chat turns.

The project's GitHub repository lists Roc Yu (@rociiu) as creator and maintainer, with two credited contributors, and showed 700 stars and 50 forks when The Terminal read it. Stars are a measure of interest; they do not say how many people have deployed it. The Terminal read the README and Cloudflare's documentation and did not deploy Talorys.

What the installer puts in your account

Setup needs Node.js 22 or later and one command, npx create-talorys@latest. It logs into Cloudflare, asks for an agent name and an owner password, then deploys a Worker and a Pages site and checks that both respond.

The README describes a React app on Cloudflare Pages that calls a Pages Function, which forwards requests over a service binding to a private Worker with no public URL. The Worker hands work to a TalorysAgent Durable Object, a Cloudflare object that holds state in SQLite and can set alarms, which Talorys uses for reminders. It does not use R2, D1, KV or Vectorize, so nothing in the stack is billed separately.

The only chat model the README names is @cf/zai-org/glm-4.7-flash on Workers AI. Semantic recall is off by default.

What 10,000 Neurons a day buys

Cloudflare's Workers AI pricing page gives every account 10,000 Neurons a day at no charge, reset at 00:00 UTC. Above that it charges $0.011 per 1,000 Neurons on the Workers Paid plan. For glm-4.7-flash the page lists $0.060 per million input tokens and $0.400 per million output tokens, equal to 5,500 Neurons per million input tokens and 36,400 per million output tokens.

The arithmetic below is The Terminal's, applied to those rates. A turn of 2,000 input tokens (the conversation so far) and 500 output tokens costs 11 plus 18.2, or 29.2 Neurons, so 10,000 Neurons covers about 342 such turns. Longer histories push the input side up, and output tokens cost more than six times as much per token.

Usage pattern	Neurons per turn	Turns per day on 10,000
2,000 tokens in, 500 out	29.2	about 342
8,000 tokens in, 500 out	62.2	about 160
500 tokens in, 200 out	10.0	about 1,000

Talorys says that when the allowance runs out, chat pauses until the daily reset while the other features keep working. The pricing page does not say whether free Neurons apply to every model, so the table assumes they do.

The Durable Objects side has its own ceiling. Cloudflare's documentation lists the Workers Free limits for SQLite-backed objects as 100,000 requests, 5 million rows read and 100,000 rows written per day, with 5 GB of stored data. A single user is unlikely to reach those figures before the Neuron cap bites.

Security claims and limits

The README says the owner password is hashed locally with PBKDF2-SHA256 and stored as a Cloudflare secret, a 256-bit session secret is generated, and authentication happens in the Worker rather than the frontend. It also says Talorys has no telemetry, and that Cloudflare still processes chat messages for inference. These are the project's claims; The Terminal did not audit the code.

The project is single-user, with no accounts or teams. New pages.dev sites can take a few minutes to pass health checks, and the installer needs specific Workers, Pages and Workers AI permissions on the token it uses.

Other Cloudflare developer stories on the site include the Deno runtime deploy deadline and the Clef decision models. The number to watch for Talorys is how many people hit the daily Neuron limit and move to the $0.011-per-1,000 paid rate.