Helmsman runs Claude Code and Codex in the same live terminal grid. You can install either CLI on its own. Codex support requires version 0.153.4 or later; native integration evidence uses 0.153.4. Support for the other agent CLIs remains available, with their own capability limits.
These additions shipped in Helmsman 0.13.0-beta.1. A Windows install stays on 0.12.0-beta.1, which predates them.
Available for Claude Code and Codex
- Launch settings. Choose the provider and native model. Codex has separate reasoning, sandbox and approval settings. Claude permission modes keep their own meaning.
- Instructions. Edit provider-specific global or project guidance. Codex supports AGENTS.md and its override file. Conflict checks and backups protect existing edits.
- Skills and plugins. Select the provider in the library or marketplace. Codex uses native skill discovery and plugin commands. A Claude plugin is not automatically a Codex plugin.
- Helmsman Apps. Provider-specific installs keep skills, runtime and app data separate.
- Usage. Session tokens, allowance observations, history and Profile statistics keep provider context. Codex dollar amounts are labeled API-price estimates when available, not subscription bills. Unavailable values remain unknown.
- Work and Fleet. Codex can discover Fleet MCP, start workers and read results. Work retains its human Accept step; an agent completion does not accept its own work.
Current Codex limits
Automatic Fleet follow-up, steer and interrupt remain disabled. Use the visible Codex terminal for these interactions. Its native controls operate on that cell’s conversation. Plan mode also uses the native terminal. Experimental App Server control is not part of the shipped adapter.
The instruction editor does not show every effective nested or managed instruction source. Skill acquisition and managed/plugin scope coverage remain incomplete. Plugin authentication health is shown as unknown where the native CLI supplies no verified status.
Cold crash recovery, compressed-conversation resume and every worktree restore case are not yet fully verified. A missing conversation must not be treated as a successful resume.
Built-in AI helpers
Choose OpenAI, Anthropic or the supported local-model route in settings. OpenAI API helpers need an OpenAI API key. Codex or ChatGPT login does not supply API billing. Anthropic uses its separate API key. Live OpenAI API acceptance remains pending; provider HTTP contracts and settings are tested.
These improvements expand Codex support. They do not claim complete feature parity.