Hermes Agent with Claude: setup, models and costs

Four ways to connect Claude

• An Anthropic API key. Hermes has a built-in anthropic provider that talks to Anthropic directly. You pay per token on Anthropic’s standard API prices, separate from any Claude chat subscription. This route gets Hermes’ full Claude support, including prompt caching and extended thinking.
• OpenRouter. One OpenRouter key reaches Claude and hundreds of other models, so you can try Claude for a week and switch to something cheaper without a new account. Model names look like anthropic/claude-sonnet-5.5.
• Nous Portal. Nous Research’s own subscription also carries Claude models, so one Nous account covers it.
• A Claude Max plan, with care. Hermes can sign in to a Claude Max account, but only the extra usage credits you buy on top of Max are used; the plan’s base allowance is not. Claude Pro cannot use this route at all. The Hermes docs call subscription billing “the single most common source of billing surprises”, so for most people an API key is the clearer choice.

There is also an experimental official plugin that runs Hermes turns through the Claude Code command-line tool on a Pro or Max subscription. Its own notes describe it as a review build, not a production-ready one.

Setting it up on your own machine

If you run Hermes yourself, the setup wizard handles every route. Run hermes model in a terminal, choose Anthropic (or OpenRouter), and paste your key or sign in. To do it by hand with an API key instead:

export ANTHROPIC_API_KEY=sk-ant-...
hermes chat --provider anthropic --model claude-sonnet-5-5

To make Claude the permanent default, set it in the config file under model (provider anthropic, default model of your choice). Inside a conversation, /model switches between providers you have already set up, for example /model anthropic:claude-sonnet-5-5. Adding a new provider or signing in has to happen in the wizard, not mid-chat. The wizard lists the exact model names your key can use, straight from Anthropic.

One cost trap: /model ... --once runs a single turn on another model, but it breaks Claude’s prompt cache on the way there and back, so a long conversation pays the full input price again. Switch for good, or not at all.

Which Claude model for which job

Anthropic’s prices per million tokens (input / output), standard tier:

• Haiku 5.5: $0.10 / $0.50 (prompts up to 100K tokens). Fast and cheap. Good for scheduled checks, sorting email, short summaries and anything that runs many times a day.
• Sonnet 5.5: $2 / $10. The sensible default: careful writing, multi-step tasks with tools, research with sources.
• Opus 5.5: $4 / $20. For hard reasoning and long, tricky jobs. Hermes’ own automatic model choice deliberately avoids Opus and Fable as defaults because of their price.
• Fable 5.1: $10 / $50. Anthropic’s top model. Rarely worth it for routine agent work.

A rough feel for the numbers: an agent turn that reads 20,000 tokens of context and writes 1,000 tokens costs about 5 cents on Sonnet 5.5 (20,000 × $2 plus 1,000 × $10, per million) and well under 1 cent on Haiku 5.5. Agent turns re-read a lot of the same context, which is where prompt caching helps: Hermes marks the stable parts of each request so Anthropic can bill repeat reads at the much lower cache price (on Sonnet 5.5, $0.10 per million cached input tokens versus $2 fresh).

A practical split: Sonnet as the everyday model, and Haiku for the scheduled jobs that run every hour.

Claude as the model vs Claude Code as a tool

These are often mixed up. Choosing Claude as the model changes how Hermes thinks for every task. Claude Code is Anthropic’s separate coding tool, and Hermes ships a claude-code skill that hands programming jobs to it: Hermes starts the claude command in its terminal, gives it the task and reports back. That needs Claude Code installed and signed in on the same machine, and it works whichever model Hermes itself runs on.

For most business use you only need the first: Claude as the model. The second matters if you want Hermes to build or fix software for you.

Using Claude on ReadyHermes

On ReadyHermes your agent starts on the included AI (DeepSeek V4.1 Flash, $1 of usage a month on every paid plan), so you need no key at all to begin. On a paid plan you can switch it to Claude from the workspace:

1. Open the workspace and choose the Model key tab.
2. Pick the provider and paste your key. We recommend OpenRouter with a Claude model such as anthropic/claude-sonnet-5.5: OpenRouter is the key route we have tested end to end.
3. Press Load models to see what your key can use, then pick a model.
4. Press Check and switch. We first send one tiny test message with your key, so a typo cannot leave your agent silent, then the agent restarts on your key, which takes about a minute.

You can also choose Anthropic and paste an Anthropic API key. That route goes through Anthropic’s OpenAI-compatible endpoint, which Anthropic describes as mainly meant for testing and comparing models, so OpenRouter is the steadier choice. Claude subscriptions (Pro or Max sign-in) are not available on the hosted plan.

Your provider bills you directly and we add no AI limit on your own key. To go back, press Use included AI. The free trial always runs on the included AI. New to Hermes? Start with what Hermes Agent is, or follow the step-by-step guide.