Talorys: Your Personal AI Agent on Cloudflare

⏱️ Reading time: 14 min

Setting up an artificial intelligence assistant that remembers your tasks, your notes and your to-dos no longer requires paying a monthly subscription or trusting your conversations to a third party. Talorys installs a complete personal AI agent inside your own Cloudflare account with a single command, with no external database to contract and without any developer on the project seeing a single line of what you write.

📑 En este artículo

1. TL;DR
2. What is a personal AI agent?
3. Why running your own AI assistant matters
4. How Talorys works under the hood
5. Practical examples and how to get started
   1. Step 1: run the installer
   2. Step 2: what the installer does while it runs
   3. Step 3: non-interactive installation
   4. How to confirm it deployed correctly
6. Real use cases for a self-hosted assistant
7. Common mistakes and best practices
8. Comparison: Talorys versus other ways to have an AI agent
9. Going deeper: the technical design behind Talorys
10. Frequently Asked Questions
    1. Do I need to know how to use Wrangler to install Talorys?
    2. Does Talorys keep working if the Workers AI free quota runs out?
    3. Can I use Talorys as my team’s single personal AI agent?
    4. Does Talorys send my data to any server run by the project’s developers?
    5. What happens if the Talorys installation gets interrupted halfway through?
11. References

The project is open source and runs on free Cloudflare pieces: Workers, Durable Objects and Workers AI. The promise is that a single person, with no infrastructure team behind them, can have it up and running in minutes with npx create-talorys@latest.

TL;DR

• Talorys installs a personal AI assistant on your Cloudflare account with a single command: npx create-talorys@latest.
• Chat, memory, tasks, notes and reminders run inside a single Durable Object with its own SQLite.
• The Worker is deployed with no public URL (workers_dev: false): nobody outside your frontend can call it directly.
• Workers AI runs the @cf/zai-org/glm-4.7-flash model for chat, with a daily free quota.
• Automations use Durable Object alarms, so reminders fire without keeping anything running.

What is a personal AI agent?

A personal AI agent is a software assistant that runs under your own control, not on a third-party provider’s server: it stores conversations, memory, tasks and reminders on your own infrastructure and only uses the language model to reason about that data, without anyone else seeing it.

Talorys is not a SaaS product you rent, but an installer that runs entirely inside your own Cloudflare account. Instead of paying a subscription to a third party, the project packages the entire stack (frontend, backend and database) using Cloudflare resources already available on the free plan, and that installer creates, configures and connects them without manual intervention.

It is, moreover, a single-user assistant: there’s no sign-up, no team accounts, no multi-user admin panels. The installer asks for an owner password, and that’s enough to log in.

Why running your own AI assistant matters

The value proposition of a personal AI agent is simple: your data never leaves your Cloudflare account. No external developer manages the server, and the project explicitly states there is no telemetry and no account operated by the Talorys authors.

The second reason is cost. Cloudflare Pages, Workers, Durable Objects and Workers AI have a free plan, so the entire stack can run without a credit card as long as usage stays within those limits. If the daily Workers AI quota runs out, chat stops responding, but tasks, notes, memory and reminders keep working because they don’t depend on the language model.

The third reason is operational. Talorys doesn’t need an external cron or a process that stays running: reminders and routines are scheduled as Durable Object alarms, so the agent can sit idle for hours and still fire notifications at the exact moment. Compared to maintaining your own VPS, there are no OS patches to apply and no manual backups to schedule.

Finally, the project works as a practical introduction to the Cloudflare Agents SDK. Anyone who wants to understand how an agent with persistent state is built on top of Durable Objects has in Talorys a complete, auditable example, not a toy tutorial.

How Talorys works under the hood

The architecture has four pieces. The browser only talks over HTTPS to the Cloudflare Pages site. Routes starting with /api are handled by a Pages Function, which forwards the request to a private Worker via a service binding, without exposing any public URL for that Worker. The Worker uses Hono as its router and calls getAgentByName("personal-agent") to get the corresponding instance. That instance is TalorysAgent, a Durable Object built on the Cloudflare Agents SDK that stores everything in its own SQLite database: conversations, memory, tasks, notes, projects, automations, sessions, configuration and usage.

flowchart TD
    A["Browser"] --> B["Cloudflare Pages: app + /api function"]
    B -->|"service binding, no public URL"| C["Private Worker with Hono"]
    C --> D["TalorysAgent: Durable Object"]
    D --> E[("SQLite: chats, tasks, notes, memory")]
    D --> F["Workers AI: glm-4.7-flash"]
    D --> G["Alarms: reminders and routines"]

Authentication and authorization happen in the Worker, never in the frontend. That matters because an attacker who manages to read the Pages site’s HTML or JavaScript won’t find any permission logic to exploit: everything sensitive lives on the other side of the service binding.

When you write a message, the response arrives as Server-Sent Events end to end. The model starts generating text and the browser displays it token by token, without waiting for the full response to finish.

sequenceDiagram
    participant U as User
    participant P as Cloudflare Pages
    participant W as Private Worker
    participant D as TalorysAgent
    U->>P: writes a message in the chat
    P->>W: POST /api/chat via service binding
    W->>D: getAgentByName personal-agent
    D->>D: reads memory and tasks in SQLite
    D-->>W: responds in SSE streaming
    W-->>P: forwards the stream
    P-->>U: shows the response token by token

Memory isn’t sent to the model whole on every turn. Talorys stores durable facts and preferences, but only sends the model the memories most relevant to the current message, which keeps the context short and the response cheaper.

Practical examples and how to get started

The simplest use of Talorys isn’t technical: it’s sending it a chat message so it handles something for you. This is what a typical conversation to create a task looks like, adapted from an example in the project’s official README:

User: Add a task to review my project tomorrow
Talorys: Task created: "Review my project", due: tomorrow.

That same task remains visible and editable from the UI, with full CRUD, not just from the chat. The second, more advanced example is the full deployment of the agent.

Before installing you need two things: Node.js 20.18 or a later version, and a free Cloudflare account. The installer ships its own bundled Wrangler, so there’s no need to install it separately; if you already have a global one, it detects it but doesn’t require it.

node -v
# v20.18.0 or higher

Confirm your Node version with that command before continuing: if it returns a lower number, the installer will fail the environment check.

Step 1: run the installer

The command is the same on Windows, macOS and Linux. Open a terminal (PowerShell, Terminal or your preferred shell), create an empty folder and run:

npx create-talorys@latest

This does four things: it checks your Node version, checks whether you already have a Wrangler session or opens the Cloudflare authorization page in the browser, lets you choose an account and agent name, and asks for an owner password with hidden input and a strength check.

Step 2: what the installer does while it runs

After the password, Talorys generates a 256-bit session secret, creates unique resource names following the pattern talorys-<id>-agent and talorys-<id>-web, deploys the private Worker (which creates its Durable Object with SQLite), stores the secrets, and creates and deploys the Pages project with the service binding already connected.

The expected output on the last line of the terminal is the actual URL Cloudflare assigned to the project, following that same naming pattern: something like https://talorys-ab12cd-web.pages.dev. That’s the URL you open to use your agent.

Step 3: non-interactive installation

To deploy from a script or a CI pipeline without anyone typing anything, Talorys accepts environment variables:

TALORYS_OWNER_PASSWORD=your-secure-password npx create-talorys@latest --yes --account-id <id>

If the terminal has no active Wrangler session, also add CLOUDFLARE_API_TOKEN with a token that has the permissions detailed below. The --yes flag skips the interactive confirmations and uses the value of TALORYS_OWNER_PASSWORD instead of asking for it via hidden input.

💡 Tip: if the installer gets interrupted partway through, run it again in the same folder. Talorys reconciles what already exists: it doesn’t duplicate resources or ask you for the password again if it’s already configured.

How to confirm it deployed correctly

There’s no direct way to verify from the outside that the Worker came up correctly, because it’s deployed with workers_dev: false and preview_urls: false and exposes no public URL of its own. The only visible surface is your Pages site: if the login page loads and rejects incorrect credentials, the Worker is responding on the other side of the service binding. The installer itself already runs that check at the end of the deployment, without spending a single AI model call.

Real use cases for a self-hosted assistant

The most direct case is the one-off reminder: asking via chat something like “remind me tomorrow at 9 that I need to review the pull request” and having the notification land in Talorys’ in-app center without leaving any terminal open. The alarm gets scheduled on the Durable Object and fires on its own.

The second case is the daily task summary. Talorys can send a digest of the day’s pending items as a recurring routine, so it works as a lightweight replacement for a task manager with reminders, with no subscription involved.

The third case is project memory. A developer carrying context across multiple conversations (code style preferences, architecture decisions already made, client data) can save them as durable memory and edit them from the UI when they change, instead of repeating them in every prompt.

The fourth case combines both pieces: a scheduled AI routine that reviews notes and memory on its own and leaves a summary in the notification center, without anyone having to ask for it each time.

The fifth case is purely educational: using the repository as a real example of a Durable Object with persistent state, alarms and streaming, built with the Cloudflare Agents SDK, to learn the pattern before applying it to your own project.

Common mistakes and best practices

• Outdated Node: the installer requires 20.18 or higher. If you’re on an older LTS, update it before running the command; the verification error doesn’t always make the cause clear.
• Losing the talorys/ directory: that’s where talorys.json lives, with the installation id, the account, the resource names and the URL. Without that file, a future update won’t know what already exists and may try to create duplicate resources.
• Confusing the security model: assuming the Worker has a public URL like any other Cloudflare deploy. It doesn’t: workers_dev: false and preview_urls: false are part of the design, not an oversight.
• Using the Global API Key instead of a scoped token: the standard login is Wrangler OAuth. If you use an API token instead, give it only the permissions Talorys needs, not full account access.
• Assuming the agent becomes useless without Workers AI: tasks, notes, memory and reminders keep operating even if the daily free quota runs out or the service is down; only generative chat depends on the model.

Comparison: Talorys versus other ways to have an AI agent

Going deeper: the technical design behind Talorys

The Cloudflare Agents SDK solves a problem that normally requires a separate database: how to give persistent state to a Worker, which by design is ephemeral. A Durable Object with built-in SQLite gives TalorysAgent a single instance addressable by name, with its own transactional storage that survives between invocations.

Alarms are the piece that replaces cron. Instead of an external process checking whether it’s time yet, the Durable Object schedules an alarm for a future timestamp and Cloudflare wakes it up at exactly that moment, even if it’s been idle for hours.

There’s a design trade-off worth clarifying: a Durable Object processes requests directed at that same instance sequentially. For a single-user agent this isn’t noticeable, but it’s the underlying reason Talorys isn’t designed to scale to many simultaneous users on a single instance.

📌 Note: the Worker having no public URL isn’t just a deployment choice: it changes where the attack surface lives. With no workers.dev and no preview URLs, the only entry point is the service binding from your own Pages frontend.

If you’re going to generate your own API token instead of using OAuth login, you need these exact permissions: Account – Workers Scripts – Edit, Account – Cloudflare Pages – Edit, Account – Workers AI – Read, Account – Account Settings – Read, and optionally User – User Details – Read if you want the app to show your email. Any additional permission is unnecessary for Talorys to work.

Talorys deliberately doesn’t use R2, D1, KV, Vectorize, AI Search or Workflows: all the state fits in the SQLite of a single Durable Object. That simplifies the mental model, since there’s only one place to look for any piece of data, at the cost of not being able to spread load across multiple instances.

Your next step: run npx create-talorys@latest in an empty folder with Node 20.18+ installed and send a first message to the chat to see the response arrive in streaming.

Frequently Asked Questions

Do I need to know how to use Wrangler to install Talorys?

No. The Talorys installer ships its own bundled Wrangler and handles the login, resource creation and deployment on its own. If you already have Wrangler installed globally, the project detects it but doesn’t depend on it.

Does Talorys keep working if the Workers AI free quota runs out?

Yes, partly. Tasks, notes, memory, projects and reminders don’t depend on the language model and keep operating normally. The only thing that stops is chat response generation until the quota renews.

Can I use Talorys as my team’s single personal AI agent?

It’s not designed for that. Talorys is single-user by design: there are no team accounts or roles, just an owner password. For multiple people you’d need a different project or to deploy one instance per person.

Does Talorys send my data to any server run by the project’s developers?

No. All chat, memory, tasks and reminders live in your own Cloudflare account. The project explicitly states it runs no telemetry and operates no central account or database.

What happens if the Talorys installation gets interrupted halfway through?

You run the same command again in the same folder. The installer reconciles what already exists in your Cloudflare account: it doesn’t create duplicate resources, doesn’t ask for the password again if it’s already configured, and doesn’t delete any data.

References

• GitHub: rociiu/talorys: the project’s official repository, with the source code and architecture and deployment documentation.
• Cloudflare Docs: Durable Objects: official documentation on the persistent state model used by TalorysAgent.
• Cloudflare Docs: Workers AI: documentation for the inference service that runs Talorys’ chat model.
• Cloudflare Docs: Agents SDK: documentation for the SDK that TalorysAgent is built on.
• Node.js: official site for the required runtime, version 20.18 or higher, to run the installer.

📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion

Featured image: Foto de Igor Shalyminov en Unsplash