monstersMNSTR.0A20ITStar, GitHub stars 4

Using your subscriptions: ChatGPT, Codex and Claude Agent

How to use the ChatGPT, Codex and Claude subscriptions you already pay for, and what Monsters does and does not do with them.

This guide on GitHub · Read from the repository when the site was built.

On this page
  1. ChatGPT (Plus and Pro)
  2. Codex
  3. Claude Agent
  4. For all three

If you already have ChatGPT Plus or Pro, or a Claude subscription, the monsters can work with it. Monsters uses them only in the ways OpenAI and Anthropic allow for open-source apps that run on your computer (D-005, D-035). You don’t need them to get started: free OpenRouter is enough.

Everything is done from Settings → Models → A subscription you already have (Impostazioni → Modelli in Italian). Then, in Each monster’s model, you choose which monster uses what.

ChatGPT (Plus and Pro)

  1. Press Continue with ChatGPT. The browser opens on OpenAI’s page.
  2. Sign in and approve Monsters. The browser returns to http://127.0.0.1:1455/auth/callback (or to another port, if 1455 is taken) and the settings page updates by itself.
  3. The first time, OpenAI’s message «You’re using your ChatGPT plan» appears: press Got it.
  4. Choose ChatGPT as a monster’s first model. Next to the message field you see Using ChatGPT plan and the Manage usage link, which leads to ChatGPT’s settings, where you decide how much of your plan Monsters can use.

Good to know:

  • The tokens are kept in ~/.monsters/secrets/, encrypted with a key in the system keychain (or, where there is no keychain, in a file readable only by your user). They never end up in the log.
  • If you reach the limit, the monster stops and shows you OpenAI’s message with Manage usage. Monsters never switches by itself to another paid provider: you decide whether to change model.
  • Plan usage (Settings → Models, under the plan, and in a monster’s usage): the windows of Claude (session, week, Opus) and of ChatGPT, with the percentage used and when they reset. Monsters asks the service nothing and doesn’t read the programs’ credentials: it shows what Claude Code and Codex write while a monster works, so the number updates only when a monster works on that plan (“updated 20 min ago”).
  • Disconnect revokes access. If OpenAI doesn’t respond, also disconnect Monsters in ChatGPT’s settings.

Monsters on another computer (SSH)

The browser has to come back to the computer where Monsters runs. Before pressing Continue with ChatGPT, open a tunnel from the computer with the browser:

ssh -L 1455:127.0.0.1:1455 user@monsters-computer

Keep the connection open until the sign-in is complete. If Monsters picked another port, the settings show it in the «Finish signing in, in your browser» box: use that one in the command.

Codex

Codex is OpenAI’s agent for coding and using the computer. It uses the same ChatGPT sign-in: there is no second login and Monsters never uses codex login.

  1. Connect ChatGPT first.
  2. On the Codex row, press Install and connect (or Connect if you already have the codex command). Monsters installs the official @openai/codex package 0.159.2 (Apache-2.0, about 450 MB) in ~/.monsters/harness/codex, only when you ask.
  3. Choose Codex as a monster’s first model.

Codex works in the monster’s folder (~/.monsters/workspaces/<monster>) with its own settings in ~/.monsters/harness/codex-home, separate from yours. The Monsters tools (notes, web, squad) reach it as an MCP server, with the monster’s permissions. Without the terminal permission, Codex can only read. Monsters that can use the web search with Codex’s search, included in the plan. Codex’s commands have no network access (Codex’s sandbox): publishing or sending from the terminal doesn’t work.

Claude Agent

Claude Agent uses Claude Code installed on your PC and your subscription. You sign in to Claude yourself, in the terminal: Monsters has no login button, doesn’t ask for or store tokens and doesn’t read Claude Code’s files.

  1. On the Claude row, press Connect.
    • If you already have the claude command, Monsters uses it.
    • If you don’t, you can press Install and connect: Monsters installs the official component @anthropic-ai/claude-agent-sdk 0.3.286 in ~/.monsters/harness/claude-agent. It has a proprietary Anthropic license, which is why it isn’t bundled with Monsters.
  2. If the settings tell you to, open a terminal, run claude and sign in with your subscription (/login). Then press Check again: Monsters asks Claude Code whether you’re signed in (claude auth status).
  3. Choose Claude Agent as a monster’s first model (sonnet, opus or haiku).

Good to know:

  • Monsters removes ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN from the Claude Code process: with an API key you would pay as you go instead of using the subscription. It also removes Monsters’ own keys and settings.
  • Claude Code works in the monster’s folder, where Monsters writes AGENTS.md and a CLAUDE.md that imports it. The Monsters tools reach it as an MCP server; it uses its own commands and files only if the monster has the terminal permission.
  • If the settings say Claude Code would use an API key, sign in with claude and your plan.
  • Monsters that can use the web search with Claude’s search (WebSearch), included in the plan; they read pages with Monsters’ own tool, which treats them as data and not as instructions.
  • The brake applies here too: before a terminal command, a search or a service connected through Claude, Claude Code asks Monsters. If the action is one you want to decide yourself (publishing, writing to third parties, spending, deleting for good), the card “Rhea wants to …. Proceed?” appears and the monster waits for your answer.

Services through Claude

Slack, Gmail, Drive and Calendar, Figma, Canva, Asana, HubSpot, Vercel and Claude Design only accept approved apps: Claude is one. With the Claude subscription connected, in Settings → Connections → Apps and services (Impostazioni → Collegamenti → App e servizi) these services say Connect through Claude. The monsters that work with Claude use them.

  1. Connect the service to your Claude account, if you haven’t already: on claude.ai, Settings → Connectors (the panel has the link). For Claude Design, instead, open a terminal on this computer, type claude and then /design-login.
  2. Go back to Monsters and press Connect: Monsters asks Claude Code whether the service is ready.
  3. If Claude asks you to sign in again, reconnect it from Claude’s connectors, or in a terminal: claude mcp login "claude.ai Figma" (with --no-browser if you’re connected over SSH). Then Check again.

Monsters never sees these services’ passwords or tokens and doesn’t change Claude Code’s configuration: it only remembers which services you chose and which monsters use them. Disconnect removes them from the monsters; they stay connected to your Claude account until you remove them on claude.ai.

For all three

  • Every step ends up in the action log; the Stop everything button also closes Codex and Claude Code at once.
  • The limits of every job (steps and tokens, the limits settings) apply to them too.
  • OpenAI’s and Anthropic’s rules change often: we recheck them at every release that touches the providers.