The Problem
A coding agent that can edit files and run commands is only useful if it can act without being watched. But it cannot act entirely unwatched — sooner or later it needs a decision that is not its to make. Delete this. Force-push that. Answer a question about which of three approaches you actually wanted.
Run one agent in a terminal and you sit with it, and that works. Run twenty and the model breaks down. You cannot keep twenty terminals in view, you cannot tell which one is blocked, and the transcript of what happened disappears the moment you close the window.
The Approach
Exa Fleet puts every session on every machine into one browser tab, and takes three positions about how that should work.
The request should arrive structured, not scraped. The machine drives the agent programmatically, so an approval can be rendered as an actual diff and a question as actual options — rather than as text pulled off a terminal screen.
The console should not be able to reach your machines. It issues them their identity, and holds no key that opens them. The machines dial out. A break-in on the web tier does not become root across the fleet.
Safety should fail towards caution. Tool calls are graded independently at both ends and the stricter grading wins; an unparseable command is treated as destructive; an unknown tool is never assumed to be read-only.
What It Is Not
It is not a model, and it is not a wrapper around one. It runs the official Claude Code CLI on your machines and holds no API key of its own — the machines authenticate the agent themselves.
Get In Touch
The quickest way to reach us is the contact page.