Talorys: a personal AI agent on Cloudflare's free tier hits 160 points on HN

Talorys is a personal AI agent that users deploy themselves on Cloudflare's free tier. The project appeared on Hacker News on October 10, 2026, and had collected 160 points and 86 comments by the time of publication. The post is a single link to the rociiu/talorys repository, with no description. Everything below therefore relies on what is visible from the title and Cloudflare's public terms. Implementation details need to be verified in the code.

Why Cloudflare's free tier

The main challenge for a personal agent is that it has to run around the clock. It receives messages, runs on a schedule, and remembers context. The usual solution is to rent a VPS or keep a home server, which costs either money or administration time.

Cloudflare covers a large share of these needs in one place. According to its public documentation at the time of writing, the free Workers tier allows about 100 000 requests per day. Alongside it sit storage options: KV, D1 (built on SQLite), and R2 for files, plus Cron Triggers for scheduled tasks. For a single user, that is more than enough. Cloudflare changes its exact limits from time to time, so check them against the current pricing page before you launch.

There is also a less obvious detail. The 10 ms limit per invocation applies to CPU time only, not to time spent waiting. An agent that mostly waits for a response from an external language model fits within these limits better than it might seem at first glance.

Where the limits are

The term self-hosted comes with a caveat here. The code and data belong to the user, but everything runs on Cloudflare's infrastructure. This is not a home server in a closet, and full autonomy is not on offer. On the other hand, it is not someone else's SaaS either, where you control neither the logic nor the storage.

Several practical questions remain that the post does not answer:

• which model powers the agent: Workers AI or an external API with your own key;
• where memory and conversation history are stored;
• which tools and integrations the agent can call;
• how authorization works, so that a stranger cannot send messages to your agent.

The last point matters most. An agent with access to your email, calendar, or files becomes an attack target if its endpoint is open to everyone.

What we know so far

• Project: Talorys, GitHub author — rociiu
• Format: a self-hosted personal AI agent
• Platform: Cloudflare's free tier
• Reception: 160 points and 86 comments on Hacker News
• Publication date: October 10, 2026
• Documentation and feature list: in the repository, not in the post

How to approach the project in practice

Start by reading the repository, not by deploying. Open the Workers configuration and check which bindings the project requests: KV, D1, R2, secrets. Check where model requests go and what data leaves your account when they do.

Next, read the Hacker News comments. In a discussion of 86 replies, real pitfalls usually surface: hit limits, trouble with long-running tasks, security concerns. That is faster than learning the hard way yourself.

If you decide to run it, use a separate account, or at least a separate Cloudflare project. Do not give the agent access to work email and documents until you are sure the entry point is secured. Turn on usage alerts so that a runaway loop does not eat your daily limit. And decide in advance what happens when the limit runs out: most free tiers simply start rejecting requests rather than sending a bill. For a personal assistant, that is more of a plus.