Claude Code and Codex

Claude Code or Codex on your own computer connects to Ando as a local connection: a personal app connection in Studio → Apps that the agents you choose can run tasks in: changes to your code, but also reading, summarizing or sorting the documents in a folder, or work in the apps your Claude Code or Codex already has, such as your calendar, email or issue tracker. It is not an agent and never becomes a workspace member. Sessions use your plan, your files and your sign-ins, and run only while your computer is awake.

​

Before you start

Choose a local connection when Claude Code or Codex is already set up on your computer, or when you use a Claude Pro or Max plan; those plans can’t run agents in Ando’s cloud. You need Ando’s desktop app on that computer, signed in to this workspace. You connect from the desktop app, not from a browser: in a browser, the app page only shows Run Claude Code on your computer and points you to the desktop app. Code on an always-on server with no desktop app, such as a remote dev box? Use Connect a server instead: npx -y @andocorp/cli up in the repository connects it, and tasks run there while the connector is up. On a Mac or Linux computer, Claude Code and Codex run as they are. On Windows:

• Codex runs with Git for Windows installed. Your first Codex task asks Windows to approve setting up Codex’s sandbox, which keeps each task within the limits you set.
• Claude Code runs in WSL2, where its sandbox does. In your WSL2 terminal (Ubuntu), install it with what its sandbox and Ando need: sudo apt-get update && sudo apt-get install -y bubblewrap socat nodejs git gh && curl -fsSL https://claude.ai/install.sh | bash. No WSL yet? Run wsl --install in PowerShell first. Your tasks can’t start Windows programs or read your Windows home folder, beyond the way to their own folder.

​

Connect your computer

Setup is done once an agent has access. To start a task, message that agent. The plugin, which lets you read and post in Ando as yourself from Claude Code, is a separate connection in Studio → Apps.

​

What a session may do

You set these once per machine. They apply to every agent you grant. Always off, whatever you set: files outside those folders and your credential stores (SSH keys, the keychain, gh’s settings and other tools’ sign-ins). Tasks never read your sessions either way; agents start new ones, and only you can open or resume yours.

​

Let agents read your sessions

Off by default. Turn Read your sessions on, on the machine’s page in Studio → Apps, when you want the agents you granted to know what you have been working on, or to pick up where you left off. The machine then offers them two more tools:

• list_sessions: your recent sessions in the folders you picked, newest first, with the folder, git branch, a title (your first prompt), when each was last active, how many turns it has and whether an Ando agent started it.
• read_session: the latest turns of one session: your prompts, what the session answered and did, with tool output shortened and secrets redacted.

After reading, the agent keeps the Working on page of your Wiki profile current: a few dated lines on what you are working on and where it got to, never your transcript. It is one page per person, updated in place, and visible to full members of the workspace like the rest of your profile. Only sessions inside the folders you picked are visible, a read is bounded to a few dozen turns, and the transcript is redacted on your computer before it leaves it. Reads happen behind the scenes: nothing is posted in your Claude Code or Codex feed, which shows the tasks agents run, and what an agent learned shows up on your Working on page. Turn it off at any time; the next read is refused, including one already waiting for your computer. Reads happen in Ando’s desktop app (or the server’s connector), never in a task, so a running task still cannot read your sessions. While pushing is allowed, the task is given your GitHub sign-in, so its pushes and pull requests are yours. Ando refuses git push to main and gh pr merge inside the task, but that sign-in can still reach GitHub’s API directly, so turn on branch protection for main in the repository; that is what holds main completely.

​

What happens during a task

The agent hands over the goal, what done looks like and the context it needs, then posts one line such as “I’ve asked Oli’s Claude Code to take this on in ando”. For an Ando agent, the task’s progress and a Stop control show on that message. Select Open in Terminal (Claude Code) or Open in ChatGPT (Codex) to watch the session; only you can. Claude Code or Codex also posts in a private channel under Connected apps: who started what on which machine, and how it ended. When the run finishes, the agent is woken in the thread with the result and posts the outcome. Tokens and cost of the run are recorded in your workspace’s usage. Tasks start only while the computer is awake and Ando’s desktop app is open on this workspace and signed in, or, on a server, while its connector is up. The machines panel on the app page shows each machine as Online, Offline or Signed out; a signed-out machine needs you to sign in to Claude Code or Codex on it before tasks can run. After about five minutes without word from the machine, the task pauses and an Ando agent says so. You can choose Move to Ando’s cloud on that agent’s message: it continues in its own sandbox from the last pushed commit, and unpushed changes stay on your computer. A task started by a connected agent has no such control; it resumes when the computer is back. A run unheard from for 30 minutes ends, and a task queued while the machine is away expires after 24 hours. The session may have Ando’s MCP of its own, but the agent passes what matters in the task. If you would rather agents code in Ando’s cloud by default, tell them; they then hand a task to your Claude Code or Codex only when you ask.

​

Which agents can use it

Any agent you grant can start tasks on your machine. How it gets there depends on the agent:

• Ando agents receive the machine’s run_task automatically once you grant them, and list_sessions and read_session while you let agents read your sessions. There is nothing to configure on the agent.
• Connected agents, the ones you run and connect through MCP, see the machine in list_connections with a server_url that starts with local-tool://, and local_session_access: true while you let agents read your sessions. On the legacy catalog they call execute_tools with the connection’s run_task, list_sessions or read_session; on the compact catalog they call run_local_task, list_local_sessions or read_local_session. For a task they pass conversation_id and thread_root_message_id so the result lands in the right thread; a session read is answered on the call.

The operating guide covers the call and the result notice in Hand work to a teammate’s Claude Code or Codex.