Technical overview

The Chili Principles

A privacy-first architecture where your data, conversations, and model choice stay under your control.

Privacy model

Chili is built on a user-first data principle. Your conversations are not used to train any model. Nothing is retained beyond your active session unless you explicitly enable memory.

This model is grounded in how LLMs actually work, where each message to LLM models is a stateless forward pass, and what is different is how different LLM providers handle your data in the rest of their systems. Learn how LLM inference works →

  • No training on user conversations
  • No data retention beyond session by default
  • User-owned memory and data, you decide what persists

Persistent memory

Chili builds a persistent memory graph across your sessions, to build an interaction that is more meaningful and useful for every users.

Existing memory implementations require users to build custom packages or use other platforms, where Chili take LLM and agent memory as a core service for all users.

LLM/Agent tools

Chili enrich your interactions with various tools that come with a modern GPT/LLM service. Current and planned:

Agent tools
Web search
Web fetch
MCP server for more tools extensions
Multimodal support
PDF, images, rich text
Project sandbox folder

Model options

Chili is designed to work with multiple LLM provider, and giving you full control.

01
BYO API key
Connect your own Anthropic, OpenAI, or Deepseek API key.
02
Off the shelf providers
Access to OpenAI, Anthropic, Gemini, Deepseek APIs
03
Chili hosted
Use Chili's hosted open-weight models. Managed, no setup. Coming soon