Skip to main content
A session starts the moment you send your first message from the home screen. Alpha creates the session, boots a machine for it alone, and hands your prompt to the agent, usually before you have finished reading your own message. To choose the agent, model, setup or repositories first, set them on the home screen. To go back to a session you already started, find it in the sidebar.

What happens when you press Enter?

Starting a session takes four stages. Keep the tab open until your message has gone: attached files upload from your browser, and the message follows them. Once the agent has it, the turn runs on Alpha’s servers, so you can close the tab and the work carries on.
1

Send the first message

Press Enter in the home composer, or click the arrow button (Start the session). Everything you set on the home screen goes with the message: the agent, the model and effort, the permissions, plan mode, the setup, the repositories and any name you gave the session. Alpha creates the session and starts booting its machine at once, while your message is still on its way. While the session is being created, a second Enter does nothing, so you never get two sessions by accident.
2

Watch the session open

The home screen dissolves into the session’s layout, and the composer glides down to the bottom of the chat. Your message appears straight away as your bubble at the top of the transcript, and the header already carries the session’s name: at first, the opening of your message. If sending takes more than a second, “Sending…” shows under the bubble. Browsers without view transitions, and systems set to reduce motion, skip the animation.
A new session a few seconds after its first message, with the sidebar collapsed. The header shows the title Harbor Sailing Classes and a pill with the Codex mark reading codex, and at its right the round +, GitHub and Share buttons and the lit files button. There is no workspace panel. The files panel is open along the right edge, headed /workspace with Refresh and Hide buttons, and reads workspace is empty. At the top of the chat, the prompt asking for a class-booking web app for Harbor Sailing School sits in a bubble. Below it, the agent's first message reads I'll build the weekly class list and booking form, keep availability in server memory, and check that a booking reduces the seat count. I'm using the design-system and landing-page-design skills for the interface. Under it, a folded row with two small round icons reads 3 commands, and a three-by-three pixel loader sits beside Striking while hot… and a 4s counter. The composer at the bottom shows the placeholder Add a follow-up, the + button, the model chip reading gpt-6.1-sol · high, the plan chip, the microphone and a round Stop button with a dark square.

Your message sits at the top, with the agent's first reply and the working line under it.

3

Let your files upload

Attached files wait for the machine, then upload into /workspace/uploads/. Meanwhile a status line directly above the composer reads “Uploading 2 files…”, with a seconds counter after three seconds. Each file adds a line such as [Attached file: uploads/class-schedule.csv (12.4 KB)] to your message, so the agent knows where to look. Images up to 15 MB also reach the model as pictures it can see. Files and uploads covers attaching more later.
4

Follow the first turn

The machine boots from a ready image in about a second, with the agent already warm inside it. Repositories you chose are cloned before the turn begins; a setup starts with them in place. While the agent works, a pixel loader shows a rotating word such as “Thinking…” or “Striking while hot…” and counts the seconds. Then the transcript streams commands, edits and messages live; Follow the agent’s work explains each part. Until the machine is up, the files panel reads “Sandbox is loading…” and workspace tabs that read from the machine, such as Preview and Database, show grey placeholder bars. They fill in by themselves once it is running.
A first message that is a command, such as /init, opens the session with the command as a chip instead of a bubble. The status line reads “Starting the machine…” until the agent takes the command. With files attached, it reads “Uploading 2 files…” first, then “Handing /init to the agent…”.

How does the session get its name?

A session’s name changes twice in its first moments, then stays put until you change it. The short name replaces the 60-character one only if nobody has renamed the session in the meantime. A name you set with /rename on the home screen is never replaced. To rename a session later, use Rename in its ⋯ menu in the sidebar.
If new sessions keep the first 60 characters of their first message as their name, short names aren’t turned on for your Alpha.

Read the session header

The header runs across the top of the chat. It tells you which session you’re in, what runs it and what it has cost so far.
A close crop of the top left of a session header. A menu button for the collapsed sidebar comes first, then the title Harbor bookings, then the Codex mark on its own, without the word codex.

In a narrow chat column, the header keeps the title and the agent's mark.

When the workspace is hidden, the right of the header also holds the + menu (Add a tab), the GitHub and Share buttons, Show workspace and Show files. While the workspace is showing, press ⌘⇧B to fold the chat away and give the workspace the whole window. Press it again, or click Show chat at the left of the workspace’s tab row, to bring the chat back. Use Ctrl instead of ⌘ on Windows and Linux. Every shortcut is listed in Keyboard shortcuts.

Open a session in a new tab

Each session lives at a permanent address, /s/{code}, where the code is 16 random characters. The address bar follows the session you have open, so you can bookmark it or paste it into another tab. The same address is the session’s share link. Until you share the session, only you and your Alpha’s admins can open it. Several sessions can work at the same time, each on its own machine, so a second tab is a good way to run two jobs side by side.

What if the session doesn’t start?

A start can fail in two places. Neither loses what you typed. A message longer than 100,000 characters is refused with “prompt too large”; attach the long text as a file instead. For other messages and their fixes, see Troubleshooting and FAQ.

What limits apply from the start?

These apply from the first message. Values marked “(default)” are set by whoever runs your Alpha. The full list is on Limits.

Next steps

Follow the agent's work

Read commands, edits, plans and sub-agents as the first turn streams.

Sleep, wake and reset

What the machine does after you leave, and how to start it over.

Setups: ready-made machines

Start sessions with your repositories cloned and dependencies installed.

Queue, steer or stop

Send your second message while the first turn is still running.