Skip to content

Agents: Claude Code, Codex, Antigravity

The runner daemon can spawn three coding agents. A machine spawns only the ones it lists, and that is Claude Code alone until you add another.

Agent Command Sign in once with
Claude Code claude claude, then follow the sign-in
Codex CLI codex codex login
Antigravity CLI agy the CLI’s own sign-in

The platform never holds any of these logins. The spawned session authenticates exactly the way your own terminal does, so a machine whose agent is signed out boots perfectly and fails every run. That is why cawdev --setup asks about it.

Terminal window
cawdev config agent enable codex
cawdev config agent disable codex

Or press c in the attached terminal and pick Enable codex. Both write agentCommands in the config file, and agentCommands is hot-swappable: a running daemon picks it up with no restart.

The first of these that answers:

  1. The run’s own choice. Agent in the composer’s More pill, where an agent the selected machine does not spawn is greyed out. In the terminal, --claude, --codex or --agy on a /new or /develop; /agent codex makes it the choice for every /new after it, kept across launches in ~/.cawdev/terminal.json, and /agent default hands the choice back.
  2. The project’s default agent, under Project settings → Project.
  3. Your default agent, under Your settings → Defaults.
  4. Claude Code.

A name no machine spawns is refused. It is never quietly run as Claude Code instead. In advanced mode each window says which agent and model actually ran (Codex · GPT-5-codex).

A card can name a model, and you have a default model for each agent. A run takes the card’s model, then your default for that agent. A default set for a different agent is never carried across: a Codex run is not handed sonnet because the card or your Claude default says so, because Codex refuses --model sonnet at spawn. It gets your Codex default, or Codex’s own. The same holds for a stage that has no route of its own.

A model name cawdev cannot place (not Claude’s, not GPT or Codex’s, not Gemini’s) is passed on exactly as typed.

Neither asks permission one command at a time, so the things cawdev builds on that are not available to them:

  • Writing needs a machine that allows everything. A Codex or Antigravity run that changes code is refused before anything is spawned unless the machine accepts rules from the console (cawdev config rules on) and has Allow everything ticked under Your settings → Machines. The refusal says which of the two is missing, on the run’s transcript as well as its summary.
  • Project tool rules, the permission prompt and the shield’s command half do not apply. Those are Claude Code’s --allowedTools grammar and its approve handler. The half of the shield that takes secrets out of a result runs in the daemon, so it still applies. See Permissions and the shield.
  • Experts are not reachable, and a live follow-up prompt is not either: the process ends with its turn, and a follow-up is a resumed run.

Codex runs once per turn as codex exec --json, under the sandbox the operating system enforces:

Run Codex is spawned with
One that must change nothing: ask, audit, scope, review, plan, and the plan, verify and memory stages of any run --sandbox read-only. The shell reads and cannot write, and it needs nothing from the machine.
One that writes: code, interview, merge, release --dangerously-bypass-approvals-and-sandbox, only on a machine that allows everything. Writing runs must commit and push, which the sandbox refuses.

So a Codex plan, review or question needs neither setting, and Antigravity, which has no such sandbox, needs both for every run. Codex uses the sign-in from codex login and not your ~/.codex/config.toml (--ignore-user-config), so a model or MCP server you configured for yourself never reaches a run. The run’s token reaches cawdev’s server through the environment, never the command line, and is blanked for the model’s own commands. No usage windows are read for Codex, because it has no -p /usage.