Skip to main content
This is the ten-minute path from nothing installed to a first piece of work finished in one of your own repositories. You need a terminal, a Cortex account, and a project to point it at. Sessions run on the Cloud host by default, so a fresh install completes a turn with no extra configuration. Nothing here assumes you have used Cortex before, and each step links to the page that covers it properly.
1

Install the CLI

Then check it:
If the command is not found, the install directory is not on your PATH. Install the CLI has the exact locations and the options for pinning a version.
2

Sign in

A sign-in page opens in your browser; approve it there and the CLI continues. cortex whoami confirms the account. On a machine with no browser, use cortex login --device-auth and enter the printed code on another device. See Sign in.
3

Start a session in your project

The session view opens: a timeline above, a composer at the bottom with the placeholder Plan, search, build anything, the mode chip reading Agent at its top left, and the model chip, for example Cortex Mini 1 (medium), at its bottom right.On a first run you also get the tips panel:
Typing / opens the command list, and /models is the entry for picking a model and its reasoning effort.Now type what you want and press Enter:
Cortex works through the request as a series of tool calls, searching, reading, editing and running commands, and each one appears in the timeline as a row while it happens. The turn ends with a line like Worked for 12s.Press Esc to interrupt a turn that is heading the wrong way; the timeline marks it Stopped. While a turn is running the composer still accepts text, but Enter queues a follow-up instead of sending it.
4

Tell Cortex about the project

This writes an AGENTS.md in the current directory, and /init does the same from inside a session. Cortex reads that file at the start of every session in the project, so put the build command, the test command, house style, and anything a new contributor would need to be told into it. Add --force to overwrite an existing file.A good AGENTS.md is the single highest-value thing you can do for the quality of later turns.
5

Choose how much autonomy to grant

Every session runs under an approval policy and a sandbox. The defaults are approval_policy = "on-request", so the agent asks before anything consequential, and sandbox_mode = "workspace-write", so writes stay inside the workspace.Press Shift+Tab to cycle the mode. The chip at the top left of the composer tells you where you are:Plan and Ask are enforced by the harness, not by asking the model to behave: the mutating tools are withheld outright.Or set the policy when you start the session:
When the agent wants to run a command you get a numbered prompt headed Cortex wants to run. Option 1 runs it once, option 2 allows that kind of command in this project from now on, option 3 lets you edit the command first, and option 4 refuses and lets you say what to do instead. See Modes and permissions.

​
Other ways to start

The session view needs a real terminal on both standard input and standard output. In a pipeline or a scheduled job, use cortex run or cortex exec instead: see Headless and one-shot runs.
Every turn goes to the Cortex coding service; the CLI does not run a model on your machine. If the service cannot be reached you get The coding service is temporarily unavailable and a non-zero exit status, never a quiet fallback.

​
Where to go next