Skip to content

Hand big coding work to a coding agent

Say, "Take the refactor to Codex and work in our shop repository." Player One starts a Codex session in that folder, gives it the job, and reports back in the chat you asked in. You can watch the work in a chat of its own, and the results land in your files. Player One drives Claude Code, Codex, Gemini CLI, Kimi Code and OpenCode this way, on accounts you connect, instead of trying to replace them.

Small jobs stay in the conversation. A failing test, one file, a compile error: your agent does those itself, and a separate coding agent would only be slower. For a large or multi-file change, Player One says plainly that a dedicated coding agent will usually do it better, and asks whether to hand it to Claude Code, to Codex, or to keep it in the chat. You choose, and "keep it here" is always allowed.

This is Player One using a coding agent. For a coding agent that uses Player One, see Connect Claude Code, Codex or another tool to Player One.

What each one runs on

Each coding agent runs on its own company's models, through an account it needs.

  • Claude Code uses the Claude account you connected on Admin > Providers (/admin/providers), the one your agents use for models.
  • Codex uses your OpenAI Codex sign-in on the same page, which is your ChatGPT plan.
  • Gemini CLI uses your Google API key on the same page. It does not use a Google sign-in.
  • Kimi Code signs in to your Kimi or Moonshot account itself.
  • OpenCode uses a login of its own for whichever provider you choose.

Set one up

Each is off until you ask, and you only do this once.

  1. Say, "Set up Claude Code" (or Codex, Gemini, Kimi, OpenCode).
  2. Player One installs the coding agent on the computer your installation runs on and turns the coding-agent runtime on.
  3. It asks your permission before connecting the account, even if that account is already connected for models. Say yes, and sign in if it asks you to.
  4. It runs a short test turn and tells you whether it worked.

The first time you switch the runtime on, your installation restarts to load it. Player One schedules the restart itself and holds it back while work is running, for up to about five minutes, so the test waits for it.

You see Player One report that the coding agent is ready. There is no settings page for this: you ask, and Player One sets it up. Ask which coding agents are ready, or to test one, whenever you want to know.

Hand it a job

  1. Say what you want and which folder it works in: "Use Claude Code to add pagination to the orders page in the shop repo." If you name no folder, it normally works in your agent's own workspace.
  2. Player One starts the session and keeps it open for follow-up, unless you say it is a one-off.
  3. The coding agent works. Its final message comes back into the chat you asked in.

You see the coding agent's face in the row beside the message box, with its label or the folder it works in underneath. Open it and you read its own chat as it goes. While the session is open you can type into that chat and your message goes to the coding agent. When the session ends, because you closed it or a one-off finished, the chat becomes read-only, says "This Claude Code session has ended", and links back to the original chat.

The results are files in the workspace, under Manage > Files (/workspaces).

What it can touch

A coding agent is not sandboxed. It runs as the same user your installation runs as, on the same computer, in the folder it was given. Player One approves its requests to edit files and run commands there, so it does not stop to ask you. Point it at the folder you mean, and keep anything you would not want changed out of its reach.

Models and cost

A coding agent does not use the model rules you set for your agents. Claude Code uses Claude models only and Codex uses OpenAI models only. OpenCode uses whichever provider you gave it.

For Claude Code and Codex, which run on the subscription you signed in with, Player One picks the coding model it recommends for that company and runs it at a high level of reasoning. For Gemini CLI, which bills your Google API key by use, Player One picks no model for you and Gemini CLI uses its own default. To choose a model for any of them, ask for a routing rule for that coding agent. It then shows on Admin > Agent Models, on the Advanced Routing tab (/admin/models/routing).

What it does not do

  • It takes no attachments from the chat. Put the files in the workspace instead.
  • Its chat has no model picker, because the coding agent owns its model, and you cannot move a running session to another agent.
  • A failed turn is reported, not repeated. If it failed on an account problem, fix the account first.

If something goes wrong

It says it is not signed in. The account is expired or was revoked. Reconnect it on Admin > Providers for Claude Code, Codex or Gemini CLI, or ask Player One to sign Kimi Code or OpenCode in again, then ask for a test.

It says "ACP runtime backend is not configured". ACP is how Player One talks to a coding agent. The restart from setup has not happened yet. Wait a few minutes, or restart on Admin > Runtime (/admin/gateway), then ask again.

It reports a permission denial. Ask Player One to check the coding agent's permission setting.

Nothing happens when you ask. Ask which coding agents are ready. One that was never set up, or whose account has lapsed, will not start.

Where next