An agent on a Claude subscription: limits, a backup input and what happens when it hits one

How the agent connects to a subscription

A hub agent runs on a model you give it access to. The simplest input is the Claude subscription you already have. Under your own account, run claude setup-token and paste the token into the agent card, in the “Model access” section.

The token is write-only: once saved, nobody sees it — not you, not us — and it never reaches the agent’s memory. The subscription stays in your name and you pay the provider at its own rates. The hub does not resell compute and adds no markup.

What limits a subscription has

A Claude subscription has two limits:

• a five-hour window — how much work fits into one stretch; after the reset the window opens again;
• a weekly limit — an overall ceiling for seven days.

The exact amounts depend on the plan, and the provider changes them from time to time. The current usage is shown in the Usage section of the claude.ai settings and by the /usage command in Claude Code.

Limits apply to the whole account. If you also work on the same subscription — in the Claude chat or in Claude Code — you and the agent share one limit.

Usage depends on the work, not on the number of messages. A long conversation or a large document uses up the window faster than a short question.

What happens when the agent hits a limit

The agent learns about the limit from the provider’s response and marks the input as resting. From there, one of two things happens.

There is a backup input. The same turn is repeated on it at once: the person in the chat gets an answer, just through another input. The model on the backup input is of the same level as on the main one.

There is no backup. The agent says so plainly: the window of its model input is used up, there is no backup, and it will be back as soon as the window opens. It does not go silent, does not pretend to be busy and does not quietly switch to a weaker model. The agent retries the resting input from time to time and gets back to work as soon as the provider lets it in.

If the token has been revoked or has expired, the agent says that too: the input has to be connected again in the agent card.

What a backup input can be

In the “Model access” section you can add several inputs — a main one and backups. A backup can be:

• a second Claude subscription on another account;
• a ChatGPT subscription (Pro or Plus);
• a key from another provider — DeepSeek, GLM, Kimi, MiniMax or Qwen;
• an Anthropic API key;
• Amazon Bedrock in your AWS account, including processing in the EU only.

A request is processed by the provider whose input is active at the moment, and on that provider’s terms — worth keeping in mind when you choose a backup. Which one to keep is your decision. We do not recommend a setup; we show how much work went through each input.

How to keep an eye on limits

The cabinet shows it without any arithmetic:

• the “Subscriptions” page shows which input each agent is on, how much it used on it this week and who has hit a limit and is resting right now;
• the overview shows how many times agents hit a limit in seven days, and the agent card shows its health for the last 24 hours: answers, limits hit, revoked inputs, failures;
• “Spend” shows turns, chats and tokens by day and the cost at the provider’s published API prices. On a subscription you pay the subscription price, and this number shows the load the agent puts on it.

How to hit limits less often

• Keep a backup input. It is the difference between “slower” and “no answer”.
• Separate people and the agent. If the limit is not enough for both of you, the agent can have an account of its own.
• Move the heavy work. Large documents and long reviews can go to the agent in the morning or evening, when nobody is waiting for a quick answer.
• Watch the limits hit, not the percentages. No limits hit in a week means the headroom is enough. Limits hit every day at the same time mean it is time to change the schedule or add an input.

Where to start

The first week is free, on the subscription you already have. The agent works on real tasks, and you watch in the cabinet how much it uses and whether it hits a limit. After that you decide whether to go on and on which model input.