Urgent.News

What's breaking now, across thousands of outlets.

AI

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

Article URL: https://github.com/rociiu/talorys Comments URL: https://news.ycombinator.com/item?id=50031614 Points: 196 # Comments: 101

Talorys is a personal AI assistant designed to run entirely inside a user's Cloudflare account, without requiring any additional servers, databases, or paid services. This self-hosted AI agent can chat, remember important information, manage tasks and projects, and schedule reminders, all without needing a single command from its developers.

Talorys is an open-source project that requires only a single Cloudflare account and one command to set up. It does not provision any paid Cloudflare services, such as R2, D1, KV, Vectorize, AI Search, or Workflows. Instead, it creates a "talorys/" directory containing important installation details and deployable artifacts. This directory can be used to update the AI agent in the future without any data loss.

To use Talorys, users can install it via npm with a single command, or they can perform a non-interactive install using environment variables and a CLOUDFLARE_API_TOKEN. The browser login process uses Wrangler's standard OAuth flow, and users can manage their sessions through the Settings → Security menu. If a user forgets their password, they can simply delete the password secret in the talorys/ directory, which will sign out all devices and require a new password on the next login.

Talorys stores all data in a single SQLite-backed Durable Object (personal-agent) in the user's Cloudflare account, with no telemetry, analytics, tracking, or advertising code. All data is processed by Cloudflare's Workers AI inference service, which complies with their privacy policy and Workers AI terms. The security model and threat mitigations for Talorys are outlined in their docs/security.md file.

The AI agent has several built-in guardrails, including limits on output tokens, context tokens, tool calls, reasoning steps, AI requests, and scheduled AI runs per day. Simple reminders and task digests do not use AI. The Usage panel provides local estimates and links to the Cloudflare dashboard for exact Neuron usage details.

To start the AI agent, users can run the npm run dev command, which launches the Worker, Pages Function proxy, and Vite UI in separate terminals. A local password can be set in the .wrangler/state/apps/agent/.dev.vars file, or it can be overridden using the TALORYS_DEV_PASSWORD environment variable. The agent uses a deterministic mock AI provider, so no Cloudflare login is required for development.

Other available scripts include npm run build, npm test, and npm run test:e2e. These can be used to build the project, run tests, and perform end-to-end testing. Additionally, npm run test:pack and npm run typecheck can be used for pack testing and type checking, respectively. The npm run lint command checks the code for style and consistency issues.

To backup the AI agent's data, users can go to Settings → Privacy → Download backup, which creates a talorys-backup.json file containing all conversations, memories, tasks, notes, projects, settings, and automations. Importing this backup will redeploy the latest Worker and frontend to the same resources, while preserving the owner password, sessions, and schema migrations.

Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Read the original at github.com →

More in AI

The Hardest Problem in Personal AI Isn't the LLM

When people picture the hard part of building a personal AI assistant, they usually picture the model: which LLM, which prompt, which fine-tune.

  • The hardest challenge in personal AI isn't the LLM itself.
  • Focus on capturing information without friction and understanding relative dates.
  • Invest in retrieval pipeline, multi-platform presence, and trust, privacy, and control.

More from Saturday 10 October →