Skip to content

Choosing the agent: Claude Code, Codex or OpenCode

Every workspace's box runs one coding agent, Claude Code, Codex or OpenCode, picked when the workspace is created and kept for the life of the workspace.

Reading time
3 min read

How it works

DevBench can run several coding agents in a workspace's box: Claude Code, from Anthropic, Codex, from OpenAI, and OpenCode, which runs on your own model-provider API keys (Anthropic, OpenAI, OpenRouter or OpenCode Zen). Your organization's list is on Settings → AI, and the workspace decides which one works there. Everything around the agent works the same way for each: the terminal and conversation views, the chat and @agent, tasks, schedules, usage, and moving to another account when one runs out.

The choice is made once, when the workspace is created. A workspace's conversations, the session it resumes and the account it runs on all belong to one agent, so switching later would mean abandoning that history rather than carrying it over. The header of every workspace shows which agent it runs.

The wizard starts on your organization's default agent, which an owner sets on Settings → AI. If the agent you pick has no account ready yet, the wizard says so; you can still create the workspace, and it starts working once an owner adds an account.

The wizard also asks for the model the agent starts on. Default (recommended) lets the agent's CLI pick; an owner can set a different default per agent on Settings → AI. Unlike the agent, the model can be changed later in the workspace's settings, and applies from the agent's next session start.

Whichever agent it is, nothing in the box stops to ask anyone for permission. The container is the sandbox: approvals are off, every folder the agent works in is trusted in advance, and update offers and tips are switched off. The one step that needs a person is signing an account in.

Scheduled tasks and agent personas choose an agent too. A scheduled run starts in a fresh box, so when every account of its agent is used up it runs on another agent instead of not at all, and the run's box says why.

Steps

Following along? Tick each number off as you go.
  1. Press New workspace and fill in the first step.

  2. Under Agent, choose Claude Code, Codex or OpenCode.

    It starts on your organization's default. The choice only appears when more than one agent is available.

  3. Under Model, keep Default (recommended) or pick a model.

    Some models also offer an effort. The model can be changed later in the workspace's settings.

  4. If a note says the agent has no account ready, create the workspace anyway and ask an owner to add one on Settings → AI.

  5. Finish the wizard and wait for the box to provision.

    The chip next to the workspace's name says which agent it runs.

  6. Owners: set Default agent for new workspaces on Settings → AI.

    It only pre-selects the wizard. Existing workspaces keep the agent they have.

Worth knowing

  • The agent cannot be changed once the workspace exists. To try another one on the same work, create a second workspace on the same repositories.

  • A message sent while the agent is working steers the turn in progress, on Codex as on Claude Code. In a Codex terminal you can press Tab instead of Enter to hold your message until the turn ends.

  • Codex keeps no memory directory, so for a Codex workspace From Claude memory and devbench doc from-memory publish the workspace's instructions only.

Report a problem