PlansClaude and ChatGPT as models

What you need

• The official Claude Code CLI, or the Codex CLI, installed on the Mac and signed in to your plan.
• RoachNet 1.1. Plan models are listed only when their CLI is installed and signed in.

Turn them on

Open Settings › RoachClaw › Subscriptions and turn on Claude, ChatGPT or both. The page shows whether each CLI is installed and signed in; RoachNet asks each CLI's own status command and never reads its credentials.

Then the models appear as:

• claude:sonnet, claude:opus and claude:haiku;
• codex:<model>, for the models codex debug models lists.

They show up in RoachClaw's model list, for Org agents, and on the model server's /api/tags and /v1/models, so other tools can use them through the model server too.

How a reply is made

RoachNet runs your own CLI headless, in an empty scratch folder, with its tools off:

• Claude Code runs with no tools, no saved session and none of your hooks or MCP servers, and its text streams as it arrives.
• Codex runs in its read-only sandbox without your user config, and its answer arrives in one piece.

A client's tools and output formats are described to the plan model in words and read back from its reply, because the CLI's own tools are off. Earlier turns of a conversation are written out as a transcript.

Since 1.1.4, Let plan models use my tools in RoachClaw chat in Settings › RoachClaw › Subscriptions lets Claude Code use your skills and MCP servers in RoachClaw chat, and Codex the MCP servers, apps, plugins and skills in its config, without asking each time. Their shell and file tools stay off, and Org agents on a plan model don't get this. It's off by default.

What leaves your Mac

Your prompts go to Anthropic or OpenAI through their own CLI, under your plan, exactly as when you use the CLI yourself. RoachNet never handles your plan's credentials.

Usage and cost

Usage comes from the CLI, with cached prompt tokens counted. Each reply is traced as subscription_included and costs $0 in RoachNet; Claude Code's API-equivalent cost is kept in the trace. An Org agent on a plan model bills its runs the same way.

When something goes wrong

The error names the fix:

Related

• The Org can hire Claude Code, Codex and other CLIs as agents.
• RoachTrace shows each reply's tokens and cost.