This might sound kinda crazy, but I define “self-hosted” as something that I host myself.

For all the people complaining this isn’t self hosted, it’s open source - it’s a relatively minor tweak to replace the AI calls to point to a local model server. Everything else runs locally via wrangler. It’s less than 20 minutes of work by hand, less than a minute to ask AI to do this.

It’s a weird complaint for people who supposedly like hacker culture.

Careful with Cloudflare's "free" AI tier if you use paid workers. I was never able to get an answer on why I was billed for their confusing neuron usage for stuff that should have been covered under their usage limits. Their customer support just ignored my ticket. Apparently it's a known issue that they don't give a shit based on discussions in their support Discord.

I'm not sure I understand the use-case or target audience for this. Selfhosted on cloudflare is a great oxymoron though.

Durable Objects are the coolest thing I have seen in a while. Code and its data locked away together, frozen when not in use, but ready to go in milliseconds when a request comes in. They are so damn useful.

They had me at: "personal", which implies private when you call it self-hosted. But like this, it's nothing private, personal or really "self" at it. It's just hosted somewhere you decided, but not self and not home.

I find it troubling that people with no prior knowledge begin "self-hosting" their chatbots like this and put their secrets and personal stuff there, which this repo encourages. From a security-view it's a nightmare to advertise this.

This project falls foul of the same issues as almost every one of these personal assistants: there's no indication that it works.

Each part of this is very easy to implement, even if writing the code by hand. A chat interface, plus memories, plus semantic search, a little database, some embeddings, cron triggers. It's all trivial to implement... unless you want it to work, and more than that, if you want to know for sure that it works.

Because the harness is key, the prompting is key, the exact wording of the skills matters, the way the memories work, the embeddings used, all of these have influence over how well the system actually performs at being a personal assistant.

I've plugged together the dozen libraries necessary to do this sort of thing, and then it sorta works and sometimes doesn't. I realised that the whole problem is actually one of evaluation. If you don't have evals that say that it recalls memories with a certain performance, if you don't have evals that say that it actually completes tasks some percentage of the time, then you don't know if it actually works.

I've learnt not to bother investing my time in anything like this unless there's evidence that it actually works beyond "on my machine" from the devs.

This is the first I’m hearing about Cloudflare’s worker AI product. I see they provide 10,000 neurons a day on the free tier. What on earth is that?

Please don't use the word "self-hosted" if it's not hosted on your hardware. This thing depends entirely on cloudflares services and "generosity".

For those questioning (rightly) the "self-hosted" description, I suspect its a poorly-worded attempt to clarify there is no third-party requirement (eg. the repo is not cloudflare-owned) beyond your own cloudflare account.

Sometimes these "open source" services are a bit misleading as there is a coupling to the org providing them; I think the OP here just means there's nothing extra needed beyond whatever cloudflare commitment you already have.

There's nothing self-hosted about relying on Cloudflare for _literally_ everything.

Cool idea!

It needs semantic recall though.

To test this, I added a memory entry for "I am a vegetarian" and then opened a chat asking "give me some dinner ideas". It gave me options for chicken and salmon, among other things.

I put together a proposal to add semantic functionality.

https://github.com/rociiu/talorys/issues/2

Couldn’t seem to load the screen shot? GitHub being weird https://github.com/rociiu/talorys/blob/main/docs/screenshots...

Claude suggested the same approach for something very similar I asked and ended up deploying and using. Quite useful and snappy architecture (never used workers or durable objects). Definitely not self-hosted but whatever, I think you got the msg from all other comments.

I built it without having a chatbot interface cause I don't see much of the point, like I didn't see the improvement in usefulness 5 years ago re "natural language" task creation a la todoist.

Did a side project on Workers and the self host test for me was whether I could run it off Cloudflare at all. Durable Objects and KV bindings meant no, so it was really just deploy it yourself.

This is great. If this was available a few years ago, I would have looked at this and similar projects to see if I should base my personal use service on. Why send your sometimes relatively private info elsewhere to be profiled? Why not have software that solves for your needs?

As is, I have already been running my own service (with no intent to open source) for years where I decide features, tools and UX that suits me. My UX is substantially different from these chat boxes. It is also multi-user by design. It still needs time investment but the tradeoff is worth for me. Over time, I expect this threshold to keep getting lower and increasing percentage of folks to cross it.

Generally, as that happens with the cost and time to build software trending lower, I think it will be interesting to watch how this dynamic evolves. For example, is it possible "personal OS" systems could emerge that solve the users' needs? I have created many smallish solutions for my needs recently, but they aren't unified in any way and I personally don't have a desire to be that [human] OS.

We built a similar concept over at Ayoh.dev

We manage the app and updates, while your backend and data live in your own Cloudflare account, so it's private. It's free, and we're currently building out more apps.

I'm pretty excited by self hosted agents or maybe just self hosted MCPs to allow remote agents like Dots or Muse gated access to your data. Amazon blocks Muse/Dots but they can't block every DO or Cloudflare egress IP.

In any case, I think the generally accepted definition of self-hosting has broadened

Interesting topic for the coming months

> Your personal AI agent. Your own cloud.

Sorry, I don't own any cloud.

self-hosted, cloudflare, free == annoying captcha, in the cloud, fingerprinting, MITM, you sold all your traffic in plaintext to the bigtech blob.

Cloudflare worker AI is verrrrry slow for user interactive use cases. Fine for background jobs or evals.

Interesting subject

Neat. Gonna have to give this a try.

Isn't it a free tier until they decide you should start paying or leave?

FYI. Give Claude a cf key. Done. All work is now a prompt.

[flagged]

[flagged]

[flagged]

[flagged]

[dead]

Hi HN! I built Talorys, an open-source personal AI agent that runs entirely in your own Cloudflare account.

The idea came from wanting a personal AI assistant that I could own and control, without paying a monthly subscription or relying on someone else's hosted service.

Talorys uses Cloudflare Workers, Workers AI, Durable Objects, and Pages to provide:

AI chat with persistent memory and conversation history Task, note, and project management Scheduled reminders and recurring automations A responsive web interface accessible from any device Password-protected access with no signup or external accounts One-command deployment: npx create-talorys@latest Everything runs in your Cloudflare account. Your data stays under your control, and there's no telemetry or third-party backend operated by me.

It's designed to work within Cloudflare's free-tier limits for personal use. You don't need to manage a VPS, database server, or Docker containers.

The biggest challenge was making deployment simple enough that someone could set up their own AI assistant without manually configuring Cloudflare services.

The project is still early, and I'd love feedback from the HN community, especially on the architecture, security, and what would make a self-hosted personal AI agent genuinely useful.

GitHub: https://github.com/rociiu/talorys

It's MIT licensed. Contributions and suggestions are welcome!

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[dead]