Guide
Using a Claude Pro/Max Subscription with Hermes Agent
Short answer: not with Claude. Anthropic prohibits third-party services from routing requests through Pro or Max plan credentials, and enforces it without warning. But the thing you actually came for — a bot on a flat monthly fee, or Claude models answering your messages — still works. Here are both supported routes, and the policy in full.
Running OpenClaw instead of Hermes? See the OpenClaw Claude subscription guide.
Claude subscription auth is closed off. The two routes below are not. Pick the one that matches why you came:
Run your bot on the Plus or Pro plan you already pay for. OpenAI publishes no equivalent restriction, so this one is available today.
Set up the ChatGPT connection →The same Sonnet and Opus models, billed per token instead of per month — and for a typical bot that works out cheaper than a Pro plan. See what it costs.
Set up an Anthropic API key →What People Are Trying to Do
Hermes Agent's Anthropic provider reads a Claude Code subscription token from the CLAUDE_CODE_OAUTH_TOKEN environment variable. The idea is to set it and have Claude turns draw on subscription quota you're already paying for, instead of the pay-per-token Anthropic API. It's a reasonable thing to want, and the provider still reads the variable — which is why the question keeps coming up. Anthropic has closed it off from their side.
Why We Don't Offer a Subscription Connection
Wiring the environment variable would be the easy part. The reason we don't is that it wouldn't actually work, and the cost of trying lands on you rather than on us. Anthropic enforces this at the API layer: since January 2026, subscription tokens used outside Claude Code are rejected with “this credential is only authorized for use with Claude Code,” and there are credible reports of accounts being restricted over it. Risking a $20/mo Pro plan — or a $100–200/mo Max plan — for a connection that gets refused anyway is a bad trade, and not one we're willing to put in front of you.
This is a genuine difference between vendors, not a limitation of Hermes. Our ChatGPT subscription connection exists because OpenAI publishes no equivalent prohibition. Anthropic does, in writing, naming third-party developers specifically.
The Supported Path: An Anthropic API Key
An API key from the Claude Console is the authentication method Anthropic explicitly directs developers to. It runs the same Claude models on your Hermes instance, with no account risk and no policy ambiguity.
- Create a key at platform.claude.com and copy the
sk-ant-api03-value. - Open API Keys in your dashboard, paste it into the Anthropic API Key card, and click Save.
- Set your agent's model to a Claude model — see the table below — and message your bot to confirm Claude replies.
Full walkthrough: Hermes Agent + Anthropic API key.
If you'd rather not add a key at all, every plan includes AI credits and free models out of the box — you can deploy and start chatting without connecting anything.
What an API Key Actually Costs
Most of the coverage of Anthropic's policy change quotes figures above $1,000/month. That number describes round-the-clock autonomous coding agents, not a chat bot answering messages. Here is the arithmetic for one always-on bot at Anthropic's published list prices, assuming a conversational turn of roughly 2,000 input and 400 output tokens:
| Your usage | Claude Haiku 4.5 | Claude Sonnet 4.6 | Claude Opus 4.8 |
|---|---|---|---|
| Light — 5 messages/day | $0.60 | $1.80 | $3.00 |
| Typical — 20 messages/day | $2.40 | $7.20 | $12.00 |
| Heavy — 60 messages/day | $7.20 | $21.60 | $36.00 |
Estimates, not a quote — your own spend moves with conversation length, how much history each turn carries, and how often you reach for the larger models. The shape holds though: a typical personal bot on Sonnet lands near $7/month, which is less than a Claude Pro plan costs. Longer conversations and heavy tool use push it up; routing routine replies to Haiku pulls it down.
Available Models
| Model ID | Model | Best For |
|---|---|---|
anthropic/claude-sonnet-4.6 | Claude Sonnet 4.6 | Fast, all-purpose |
anthropic/claude-opus-4.8 | Claude Opus 4.8 | Most capable, complex reasoning |
Important Notes
- Prohibited by Anthropic: routing a hosted third-party agent through Pro/Max plan credentials is not permitted, is enforced server-side without notice, and can cost you the Claude account itself. An Anthropic API key carries none of that risk.
- Running Claude Code yourself is fine: the policy restricts third-party services acting on your behalf, not where you run Anthropic's own CLI. Using the real
claudebinary on your own machine or your own server is ordinary use. The problem is handing the credential to someone else's harness. - Subscription caps wouldn't suit an always-on bot anyway: Claude plans enforce a 5-hour rolling window plus a weekly total. An agent answering messages around the clock burns those far faster than interactive use, so even setting policy aside, per-token billing is the better fit.
Other Ways to Bring Your Own Key
- Anthropic API key — Pay-per-token, no subscription limits
- ChatGPT subscription — Use GPT-5.6 via your ChatGPT Plus/Pro plan
Or skip API keys entirely — OpenClaw Launch includes AI credits and free models out of the box.
Common Questions
I already pasted a subscription token somewhere. What should I do?
Revoke it. Sign in at claude.ai and remove the authorization, then generate a Console API key for anything automated. A token from claude setup-token is valid for about a year, so an old one left in a third-party tool keeps carrying that risk until you revoke it.
Does this affect the ChatGPT subscription connection?
No. That one remains available — OpenAI publishes no equivalent restriction on using a Plus or Pro plan through a third-party client. See the ChatGPT subscription guide.
Will an Anthropic API key cost much more?
For a typical always-on bot, it usually costs less than a subscription would — you pay for the tokens a conversation actually uses rather than a flat monthly fee. Around $7/month on Sonnet at 20 messages a day; see the full breakdown above. Claude Haiku handles routine replies for a fraction of that, and you can reserve Sonnet or Opus for harder turns.