Advanced

Use a chat plan you already pay for

This page is for you if you already pay for ChatGPT, Claude, or GitHub Copilot and want your agent to run on that plan instead of a pay-as-you-go API key. Done well, it means no second bill and no key to protect. Done carelessly, it can drain your plan in an afternoon, add charges you did not expect, or get the account suspended.

Each section says what the company allows, how billing works, and who sees your prompts. Steps are given only for routes an agent or tool officially documents.

Applies to: Agent Zero, Hermes Agent, OpenClaw, Goose, NanoClaw, and Atomic Agent have their own sign-ins. QwenPaw, Eigent, OpenManus, and Everfern have none, so the only way to use a plan there is the local tool near the end, which carries more risk.

What an account sign-in actually does

A chat plan is priced for one person chatting in the company's own app. An agent is different: one request from you can turn into dozens of model calls, because the agent plans, runs a tool, reads the result, and tries again. That is why the companies care which program is spending the plan, and why the plan's allowance can run out much faster than it does in chat.

There are three ways an agent can use your plan.

Your prompts still go to the company that made the model in all three cases. What changes is which program sends them and whether the company's terms allow it.

Which agents offer which sign-in

Agent ChatGPT plan Claude plan GitHub Copilot Other account sign-ins
Agent Zero Codex/ChatGPT account, by device code None built in GitHub Copilot account Google Cloud Gemini, xAI Grok
Hermes Agent ChatGPT or Codex Subscription in hermes model Anthropic sign-in, Claude Max with extra usage only None documented Nous Portal, SuperGrok
OpenClaw Sign in with ChatGPT (Beta), or Codex login Claude CLI, reusing a Claude Code sign-in Device sign-in OpenRouter sign-in
Goose ChatGPT Subscription, or Codex ACP Claude ACP, through Claude Code Device sign-in Cursor Agent
NanoClaw Codex or OpenCode add-ons Claude subscription is the default None documented None
Atomic Agent Through a signed-in Codex app Through a signed-in Claude Code app None documented None
QwenPaw, Eigent, OpenManus, Everfern None documented None documented None documented None

No agent here offers a consumer Gemini sign-in for new setups. That route ended on 18 June 2026, as explained in the off-limits Gemini sign-ins.

The routes side by side

Route Agents What you pay Risk What the company says
Sign in with ChatGPT, OpenAI's app sign-in OpenClaw. OpenAI also lists Hermes Agent as an app that can use your plan. Your ChatGPT plan's included allowance. Credits after the limit only if you turn that on. Safe official, while you stay inside OpenAI's rules. Surprise bill if automatic credit purchases are on. Allowed for your own requests in the app you connected. No sharing plans and no general API for other tools.
ChatGPT or Codex login inside an agent Agent Zero, Goose, Hermes Agent, OpenClaw, and NanoClaw or Atomic Agent through the Codex app The same included allowance. Surprise bill, as above. Ban risk: none reported, but no OpenAI page found here names this use. OpenAI's sign-in screen calls the access "Codex CLI".
Claude plan through Claude's own app OpenClaw, NanoClaw, Goose, Atomic Agent Plan limits while the 15 June 2026 pause holds, or usage credits if Anthropic bills it that way. Surprise bill. Ban risk if a tool disguises itself as Claude's own app. Two statements, both current. See Claude plans.
Claude plan sign-in in Hermes Agent Hermes Agent Purchased extra usage on Claude Max only. Claude Pro cannot use it. Surprise bill: every request bills as extra usage. Ban risk: Hermes says this sign-in "routes as Claude Code", and Anthropic prohibits tools that misrepresent their identity. Hermes says your included Max allowance is left untouched.
GitHub Copilot device sign-in Agent Zero, Goose, OpenClaw Copilot Free is $0. Pro is $10 for 1,500 credits, Pro+ $39 for 7,000, Max $100 for 20,000. Surprise bill once you set a budget for extra use. No agent-specific ban found. Your plan and organization decide which models you get.
CLIProxyAPI with one account of your own Any agent that accepts an OpenAI-compatible address Your plan. No proxy fee is listed. Ban risk. Surprise bill if credits or extra usage are on. OpenAI forbids general API access for other tools. Anthropic forbids routing requests through consumer plans.
Gemini CLI consumer sign-in, or Antigravity's sign-in used outside Google's app None for new setups Gemini CLI sign-in withdrawn on 18 June 2026. Antigravity's is for Google's own app only. Ban risk: Google suspended accounts in February 2026. Google's terms call third-party use a breach.

Before any sign-in: set your limits first

Do these checks before you connect an agent to a plan, because the agent will not stop on its own when the included allowance runs out.

  1. Find out whether the plan can charge you past its included allowance. ChatGPT calls this credits, Claude calls it usage credits or extra usage, and GitHub calls it a budget for additional usage.
  2. Turn that extra spending off, or set it to the most you are willing to lose in a month.
  3. Note when your allowance resets, so a stalled agent does not surprise you.

What you should see each account shows extra spending either off or capped at an amount you chose.

Note: A paused agent is a far better outcome than a surprise charge. When the allowance runs out with extra spending off, the agent simply stops working until the reset.

ChatGPT plans

OpenAI publishes two things that matter here. Its Sign in with ChatGPT Terms, effective 4 November 2026, let an app use your plan for your own requests. The same terms forbid pooling, sharing, or reselling plan usage, and using one person's plan for another person's requests. They also rule out offering "general-purpose API access for other tools or unrelated requests". Plan limits still apply, and the sign-in "does not grant extra usage".

OpenAI's help page says eligible requests "use the included ChatGPT Work and Codex allowance or applicable credits". On a personal plan, apps can use credits after the limit only if you have switched that on, and it is off by default. If automatic credit purchases are on, "continued usage can result in automatic charges," and you may not get a separate notice when included usage runs out.

Two kinds of ChatGPT sign-in

The agents use two different ChatGPT sign-ins, and the difference matters for the terms.

Both spend the same allowance. If your agent offers both, pick Sign in with ChatGPT.

Agent Zero: connect a Codex/ChatGPT account

Agent Zero's guide says you do not need an OpenAI API key or the Codex app for this, but the account needs access to Codex.

  1. In ChatGPT, open your account menu and choose Settings, then Security and login.
  2. Scroll to the bottom and turn on Enable device code sign-in for Codex, Excel, PowerPoint, and Word. On a workplace account, an administrator may have to allow it.
  3. Open Agent Zero and send a first message. When it asks for a model, click AI account, or select the Account tab if setup is already open.
  4. On the Codex/ChatGPT card, click Connect account, then click Connect account again on Choose your main model.
  5. Agent Zero opens an OpenAI page in a new tab. Go back to the Agent Zero tab and copy the Device code shown under Waiting for sign-in.
  6. On the OpenAI tab, check that the address is auth.openai.com, choose your account, paste the code, and click Continue.
  7. Wait for Signed in to Codex, then return to Agent Zero.
  8. Choose a Main model (Agent Zero suggests gpt-6.1-terra) and click Choose utility model. Keep Codex/ChatGPT Account as the Utility provider, pick gpt-6-luna if it is listed, and click Finish setup.

What you should see a green checkmark next to the account, then the Agent Zero is ready screen, and a reply to your first message in the same chat.

Note: Keep the device code private, and use only the code from the connection you just started. If the code expires, cancel the pending sign-in in Agent Zero and start again.

OpenClaw: Sign in with ChatGPT (Beta)

Run this on the computer where OpenClaw is installed:

openclaw models auth login --provider openai --method siwc

Your browser opens an OpenAI page. Approve token sharing as well as the basic sign-in, because OpenClaw's docs say a sign-in with identity permissions alone saves the account but cannot run the model. In the dashboard, the same choice is under Settings → Models → Connect provider → OpenAI → Sign in with ChatGPT (Beta), and Test & use sends a check message.

What you should see OpenClaw saves the account, and a test message gets a reply.

OpenClaw says it does not show your remaining allowance or set its own limits for this sign-in, so check usage in ChatGPT under Settings → Usage. OpenClaw also offers a Codex login (--method oauth, or --method device-code on a remote machine), which uses the same allowance.

Goose: ChatGPT Subscription

Goose's docs say this route needs an active ChatGPT Plus or Pro plan.

  1. On Goose's first-run welcome screen, choose ChatGPT Subscription.
  2. Sign in with your ChatGPT account in the browser window that opens.
  3. Authorize Goose to use your ChatGPT plan, then return to Goose.

What you should see Goose Desktop is ready for your first session. If you set Goose up earlier, the same provider is listed as ChatGPT Codex under Settings → Models → Configure providers.

Hermes Agent, NanoClaw, and Atomic Agent

Hermes Agent's own page says to run hermes model and choose ChatGPT or Codex Subscription, which signs in with a device code and uses Codex models. It also says the Hermes docs do not yet say which ChatGPT plans qualify or how Hermes usage counts against your plan. OpenAI's directory lists Hermes Agent as an app that can use your plan, but this guide could not confirm which Hermes screen offers that newer Sign in with ChatGPT option.

NanoClaw's setup can add Codex with a ChatGPT plan, or OpenCode with ChatGPT sign-in. Atomic Agent uses a Codex app that you have already signed in to. Their own guides cover the setup.

Cap what each app can spend

OpenAI's help page gives you a per-app cap for apps that use Sign in with ChatGPT on Go, Plus, and Pro plans.

  1. In ChatGPT, go to Settings → Usage.
  2. Under App limits, find the agent and select Manage limits.
  3. Set a weekly limit, as a share of your overall weekly usage, and select Save.
  4. On the same page, check the option that lets apps use credits after your limits. Leave it off unless you want the agent to keep spending.

What you should see the agent listed under App limits with the weekly cap you chose.

OpenAI says an app limit is a cap, not a separate allowance, so the agent's usage still comes out of your own weekly share. This guide could not confirm whether Codex-login routes appear under App limits in the same way.

Claude plans

Anthropic's position is the hardest to plan around, because two current Anthropic statements point in different directions. Both are quoted here because each one changes your bill.

So today, a third-party agent may come out of your plan limits, but Anthropic has reserved the right to bill it as usage credits instead, and you would pay. Anthropic's usage bundles page says bundle balance "works across Claude, Claude Code, Claude Cowork, and third-party products" and applies only after your plan limits. Bundles were $45 for $50 of credit, $200 for $250, and $700 for $1,000 when checked, with standard rates past a monthly cap.

Warning: If usage credits are turned on in your Claude account, an agent that keeps working past your plan's limits keeps spending money. This guide could not confirm the exact menu where Claude turns usage credits on or off, so check your account's billing settings before you connect an agent.

Anthropic's Usage Policy, effective 12 November 2026, forbids services "that route requests through consumer subscriptions or misrepresent the product or client being used". Claude Code's legal page says Claude plan sign-in "is designed to support ordinary use of Claude Code and other native Anthropic applications". OpenClaw's route runs Claude's own app. Hermes's route says plainly how it bills, but its docs also say it "routes as Claude Code", which is close to what the login page prohibits. Neither comes with Anthropic's written approval.

OpenClaw: reuse a Claude Code sign-in

OpenClaw can hand work to Claude Code, Anthropic's own coding app, using the sign-in Claude Code already holds. Claude Code must be installed on the same computer first.

  1. Sign in to Claude Code with this command.
claude auth login
  1. Run OpenClaw's setup and pick the Claude CLI route.
openclaw onboard
# choose: Claude CLI

What you should see OpenClaw finishes setup with Claude as the model, and a test message gets a reply.

OpenClaw says it does not copy, refresh, or store that sign-in, because Claude Code owns it. OpenClaw's docs also say Anthropic staff told the project this reuse is allowed, but that assurance was not on any Anthropic page checked for this guide. OpenClaw itself still calls an API key the safer path for production use.

Hermes Agent: Claude Max with extra usage

Hermes's providers page is unusually direct about billing. Its Anthropic sign-in "only works if you're on a Claude Max plan and have purchased extra usage credits". The same page says "Claude Pro subscribers cannot use this path". The included Max allowance is not touched at all. In the page's words, "All Hermes usage bills as extra usage even while your included Max allowance sits untouched".

Warning: Hermes's docs say this sign-in "routes as Claude Code against your Anthropic account", and that Hermes reads Claude Code's own saved sign-in when it finds one. Anthropic's login page prohibits tools that "misrepresent their identity to Anthropic's servers", so this route carries a ban risk on top of the bill. If you cannot afford to lose your Claude account, use an Anthropic API key with a spend limit instead.

  1. Run hermes model, choose Anthropic, and pick the sign-in option. You can also run hermes auth add anthropic --type oauth.
  2. Approve the sign-in in your browser.

What you should see Hermes lists Anthropic as a working provider, and replies count against your purchased extra usage.

If you do not have Max plus extra usage, Hermes tells you to use an Anthropic API key instead. A separate plugin, claude-subscription-directsdk, runs turns through Claude Code and bills your plan's Agent SDK allowance, but Hermes labels it experimental.

NanoClaw, Goose, and Atomic Agent

NanoClaw is built around Anthropic's Agent SDK, and its setup offers a Claude subscription as the default. That is the Agent SDK case that Anthropic's 7 October 2026 note covers. Goose's Claude ACP provider runs Claude Code and needs npm install -g @agentclientprotocol/claude-agent-acp plus an active Claude Code subscription. Atomic Agent uses a Claude Code app you have already signed in to. Each agent's own guide covers the setup, and both Anthropic statements above apply to all of them.

GitHub Copilot device sign-in

A device sign-in shows you a short code, and you approve it on GitHub's website. The agent then spends your Copilot plan's AI credits, so the billing works like this:

Plan Price per month Monthly AI credits Notes
Copilot Free $0 An allowance, with auto model selection only 2,000 code completions a month
Copilot Pro $10 1,500 1,000 base plus 500 flex
Copilot Pro+ $39 7,000 3,900 base plus 3,100 flex
Copilot Max $100 20,000 10,000 base plus 10,000 flex

One AI credit is $0.01. Credits reset at 00:00:00 UTC on the first day of each month and do not carry over. GitHub warns that agent-style work "can involve multiple model calls within a single task," so an agent spends credits much faster than chat does.

When credits run out, you can upgrade, set a budget for additional usage, or wait for the reset. GitHub bills additional usage "after it occurs," and you must pay for usage you ran up before you cancel or cut the budget to $0. Which terms apply depends on how you bought Copilot: personal plans fall under the GitHub Terms of Service, while Business and Enterprise plans bought directly use GitHub's Generative AI Services Terms.

Note: Leave your additional-usage budget at $0 unless you want the agent to keep spending. Copilot Free picks the model for you, so expect weaker results on hard tasks.

Goose

  1. Run goose configure, choose to configure a provider, and select GitHub Copilot.
  2. Goose copies an eight-character code to your clipboard and opens GitHub's device activation page.
  3. Paste the code and authorize the application.

What you should see GitHub Copilot listed as a provider in both Goose's terminal version and Goose Desktop.

OpenClaw

Run the sign-in in a real terminal window, because OpenClaw says it needs an interactive terminal.

openclaw models auth login-github-copilot

Visit the address it prints, enter the one-time code, and keep the terminal open until it finishes. Then pick a model your plan includes. OpenClaw's example is below.

openclaw models set github-copilot/claude-sonnet-5

What you should see the sign-in completes in the terminal, and replies come from the model you set.

OpenClaw stores the GitHub token in its local secret store, which its docs say "is not encrypted at rest". It also sends requests with a Copilot CLI request identity. No OpenClaw-specific Copilot ban showed up in the pages checked.

Agent Zero

Agent Zero lists GitHub Copilot on the same Account tab as Codex/ChatGPT. Its installation guide says the sign-in code and link appear in Agent Zero's output after your first prompt, and you approve them on GitHub's website. This guide could not confirm the other screens.

Agent Zero's other account sign-ins

Agent Zero also lists Google Cloud Gemini and xAI Grok as account sign-ins. This guide could not confirm which plan each one bills, or what Google or xAI say about third-party agents using them. Do not assume Google Cloud Gemini brings back the consumer Gemini sign-in that ended on 18 June 2026, because Google's supported route for outside agents is an AI Studio or Vertex AI key.

A local tool that passes your plan to any agent: CLIProxyAPI

CLIProxyAPI is an open-source program that runs on your computer. Its README calls it "a proxy server that provides OpenAI/Gemini/Claude/Codex/Grok compatible API interfaces for CLI". In plain terms, you sign it in to your plan, and it gives you a local address and key that any agent can treat like an API. That is how QwenPaw, Eigent, OpenManus, or any other agent with an address field could run on a ChatGPT plan.

It is also the route the companies' terms object to most clearly.

The README advertises "Multiple accounts with round-robin load balancing", meaning it takes turns between several accounts to stretch the limits. That is the account juggling this guide does not cover; add one account of your own and no more. This guide found no separate tally of bans for CLIProxyAPI users. This guide also could not confirm how CLIProxyAPI identifies itself to OpenAI.

Warning: CLIProxyAPI's documented default (host: "") listens on every network connection your computer has, not just your own machine. Anyone on the same Wi-Fi could reach it, and only the key you set stops them. In the core settings, set the host to 127.0.0.1 so only your own computer can reach it. Always set a client key, and never run it on a shared or public network.

Steps: one ChatGPT account through the desktop app

The README recommends its desktop app, EasyCLIProxyAPI, for use on a personal computer. These steps use your ChatGPT plan only, because the Claude route is the one Anthropic's policy names outright.

  1. Download EasyCLIProxyAPI for your system from its GitHub releases page: a ZIP for Windows or a DMG for a Mac.
  2. Extract the Windows ZIP and run EasyCLIProxyAPI, or open the DMG on a Mac and launch the app.
  3. Open Version Management and install the bundled or latest CLIProxyAPI core.
  4. Return to Home and start the core.
  5. Open the OAuth page, choose Codex OAuth, and sign in to your own ChatGPT account in the browser.
  6. In the core settings, add one client key: a long random password that your agent will send. Set the host to 127.0.0.1 on the same screen. This guide could not confirm the exact name of that screen.
  7. On Home, copy the OpenAI-compatible endpoint it shows.

What you should see Home shows the core running with a listening port, and your ChatGPT account appears as signed in.

CLIProxyAPI keeps your sign-in files in a .cli-proxy-api folder in your user folder, such as C:\Users\you\.cli-proxy-api on Windows. Treat that folder like a password. Mac users who prefer the terminal can install the core with Homebrew instead; the project's Quick Start has the commands.

Connect your agent to it

Point your agent at CLIProxyAPI the same way you would point it at a relay. Use the endpoint you copied, which by default is http://127.0.0.1:8317/v1, and the client key you created. Relays and aggregators has the per-agent steps for a custom address.

Two agents need a different address or have a shortcut:

What you should see your agent lists the models your ChatGPT account can use, and a test message gets a reply.

The README's sponsor section mostly advertises relay shops and residential IP services, including one pitched at "multi-account operations". You need none of them for one account of your own.

Off-limits: Gemini CLI and Antigravity consumer sign-ins

Google ended the Gemini CLI route and forbids using an Antigravity sign-in outside its own app, and this guide gives no steps for either. Google's deprecation page says that starting 18 June 2026, Gemini CLI and the Gemini Code Assist extensions stopped serving the individuals, Google AI Pro, and Google AI Ultra tiers. In Google's words, "you can no longer use the Login with Google option to access the IDE extensions or Gemini CLI".

Google's terms also forbid the practice directly. The Gemini CLI terms give "using OpenClaw with Gemini CLI OAuth" as an example of a violation. They say it "may be grounds for suspension or termination of your account". The Antigravity terms call third-party access "a breach of this Agreement".

The bans were real. On 27 February 2026, Gemini CLI developer relations posted that account problems that week came from "a series of Antigravity bans" for "the use of 3rd party tools or proxies". Because the systems were shared, those bans also blocked Gemini CLI. Google restored recently flagged accounts and said a second violation "will be permanently banned".

OpenClaw no longer creates these sign-ins, and Hermes Agent says there is no way to sign in with a consumer Gemini plan. CLIProxyAPI's README still lists Gemini CLI and Antigravity accounts, and the same ban applies there. For Gemini in an agent, use a Google AI Studio or Vertex AI key, which is the route Google names.

Trade-offs: which route makes sense

Sign in with ChatGPT in OpenClaw or Hermes Agent. Your agent spends your ChatGPT plan through the sign-in OpenAI publishes for apps.

Codex login in Agent Zero, Goose, or Hermes Agent. The agent signs in the way OpenAI's Codex app does.

A Claude plan through any agent. The agent uses your Claude plan, either through Claude Code or its own sign-in.

GitHub Copilot device sign-in. The agent spends your Copilot credits.

CLIProxyAPI. A local tool serves your plan to an agent that has no sign-in of its own.

Before you choose, run each option past the three questions in Safety versus capability.

Sources

Checked 14 November 2026.