Talorys – A self-hosted personal AI agent on Cloudflare's free tier | AI Paper Digest

TL;DR Highlight

One npx command deploys a personal AI agent to your own Cloudflare account on the free tier

Who Should Read

A full-stack developer who wants to run an agent with chat, memory, and reminders on Cloudflare's free tier without managing servers

Core Mechanics

One command deploys to your account
- `npx create-talorys@latest` creates the Worker, Pages project, and Durable Object in your Cloudflare account.
- The README says the Talorys developers operate no server, database, or account.
All data lives in one SQLite Durable Object
- Conversations, memories, tasks, notes, projects, automations, and settings all sit in one place.
- It does not use R2, D1, KV, or Vectorize.
Model is glm-4.7-flash on Workers AI
- The model is `@cf/zai-org/glm-4.7-flash` with streaming and tool calling.
- A Pages Function forwards requests over a service binding (an internal link with no public URL) to a private Worker.
Only relevant memories go to the model
- Each turn sends only the most relevant memories in the prompt.
- Semantic recall is off by default and enabled in settings or with `--semantic`.
Cost guardrails are built in

Community Reactions

Pushback: Cloudflare lock-in isn't self-hosted
- Many commenters said depending entirely on Cloudflare means it shouldn't be called self-hosted.
- One said Durable Objects and KV bindings meant it couldn't run off Cloudflare at all, so it was just deploy-it-yourself.
Opinion: it's open source, so swap in a local model
- One commenter said you can point the AI calls at a local model server and run the rest locally via wrangler.
Experience: billed despite the free AI tier
- One commenter on paid Workers said they were billed for neuron usage that should have been covered by free limits.
- They said support never answered.
Criticism: no evidence it works
- One commenter said without evals for memory recall and task completion you can't know whether it works.
- Another added a vegetarian memory and asked for dinner ideas, and got chicken and salmon.
Opinion: Workers AI is slow

How to Apply

Run it in local mock mode first
- Run `git clone`, then `npm install && npm run dev`.
- Check the UI, memory, and task structure with the mock AI and no Cloudflare login.
Build memory recall test cases
- Add 3-5 memories, ask 10 related questions, and record in a table whether they are reflected.
- Compare results with semantic recall off and on.
Start with low AI usage limits
- In Settings → AI, set low values for daily AI requests and max output tokens.
- Check neuron usage in the Usage panel before raising them.
Borrow the architecture for your service
- Design per-user agents by bundling SQLite and alarms in one Durable Object.
- Separate non-AI features from AI features so the service keeps running when quota runs out.

Try It

copy & run

# Paste into a terminal (Node.js 22+ required; run local mock mode first)

git clone https://github.com/rociiu/talorys.git
cd talorys
npm install
npm run dev

# Open: http://localhost:5173 (default password: talorys-dev-password)

# Next: real deploy uses the command below; semantic recall uses the --semantic flag
# npx create-talorys@latest --semantic

Terminology

Durable ObjectA small Cloudflare server instance that bundles code and state together. It sleeps when idle and wakes up on a request.

Workers AICloudflare's serverless model inference service. It has a daily free allocation.

NeuronThe unit Workers AI uses to measure usage. Think of it like billable minutes on a phone plan.

service bindingA direct internal call between Workers with no public URL. Like an internal office phone extension.

Semantic recallFinding memories by meaning rather than keywords. It uses embeddings.

Talorys – A self-hosted personal AI agent on Cloudflare's free tier