Skip to main content

OpenAI Codex CLI

Codex CLI is OpenAI's open-source agentic coding tool. It connects to AgentXchain via the local_cli adapter.

Before you wire Codex into a governed role, read the Authority Model. write_authority, local_cli, and Codex's own approval/sandbox mode are separate axes.

Which adapter?​

local_cli — AgentXchain spawns codex as a subprocess for each governed turn.

Prerequisites​

  • Codex CLI installed (npm install -g @openai/codex)
  • OPENAI_API_KEY set in your environment
  • agentxchain CLI installed
  • Prefer an absolute executable path when possible. On macOS app installs, /Applications/Codex.app/Contents/Resources/codex is the safest default because GUI shells and subprocess spawn contexts do not always share the same PATH.

Configuration​

{
"runtimes": {
"codex-dev": {
"type": "local_cli",
"command": ["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "{prompt}"],
"cwd": ".",
"prompt_transport": "argv"
}
},
"roles": {
"dev": {
"runtime": "codex-dev",
"mandate": "Implement features and fix bugs",
"write_authority": "authoritative"
}
}
}

Key fields​

FieldValueWhy
command["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "{prompt}"]exec is Codex's non-interactive entrypoint. --dangerously-bypass-approvals-and-sandbox is the full-authority local CLI mode for unattended governed writes.
prompt_transport"argv"The prompt is passed through the {prompt} placeholder as a command-line argument.

Authority mapping​

  • write_authority: use authoritative when Codex is expected to write directly into the repo.
  • Runtime type: local_cli.
  • Downstream CLI authority: --dangerously-bypass-approvals-and-sandbox is the required Codex mode for unattended authoritative local writes.
  • Non-interactive entrypoint: use codex exec, not top-level codex.

Absolute-path recommendation​

If agentxchain doctor says bare codex is not resolvable in the dispatch spawn context, replace the command with the real executable path instead of relying on PATH repair later:

"command": ["/Applications/Codex.app/Contents/Resources/codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "{prompt}"]

Verify the connection​

agentxchain connector check
agentxchain connector validate codex-dev

Minimal working example​

agentxchain init --governed --template cli-tool --goal "Build a JSON schema validator" --dir my-project -y
cd my-project
agentxchain doctor
agentxchain connector check
agentxchain connector validate codex-dev
agentxchain run

Or use the guided interactive path (prompts for template, name, goal, and folder):

agentxchain init --governed

Then update agentxchain.json with the Codex runtime above, run agentxchain doctor, agentxchain connector check, and agentxchain connector validate codex-dev, then start agentxchain run.

Gotchas​

  • Do not rely on --full-auto for authoritative unattended runs. Use --dangerously-bypass-approvals-and-sandbox when the governed role is meant to have real local write authority.
  • Do not use --quiet in Codex governed commands. On the current Codex CLI, --quiet is rejected; codex exec is the correct non-interactive path.
  • Model selection: Codex CLI defaults to a Codex-optimized model. You can override with --model gpt-5.4 if needed, but the default is usually better for coding tasks.
  • Working directory: Codex CLI operates on the current working directory. The cwd field ensures it runs in your project root.