Skip to content

How Felix works

You do not need any of this to use Felix, but it helps to know where things happen.

Each user gets one virtual machine with its own disk, isolated from other users’ machines. It can reach the internet, but not Felix’s own servers or other users. The machine stays on; inside it, each agent has its own container with a desktop, a Chromium browser, a terminal and a home folder. A container sleeps after 15 idle minutes and wakes when it is needed. All containers share the /workspace folder.

Each agent is a loop driven by Mistral Large 4: the model reads your message, the conversation and the agent’s instructions, then chooses an action (open a page, run a command, write a file, call an app) and sees its result, until it can answer you. The browser is driven through Playwright, the same browser you watch on the screen.

Every action passes a gate first: actions that change things outside the computer stop for your approval. Tool output is checked for your credentials, which are hidden from what the agent reads.

Your Mistral and Composio keys never enter your computer: Felix’s servers make those calls on your behalf. Built-in apps go through Mistral’s connectors with your Mistral key, and more apps through Composio with your Composio key.

Messages between agents go through Felix, never directly between containers. The master agent delegates by sending a task; the result comes back the same way. Skills are added to an agent’s instructions for the task that names them, and routines are started by Felix’s scheduler or by a call to their webhook URL.