#Feature Request

1 messages · Page 1 of 1 (latest)

silver sentinel
#

I opened this yesterday. Not much of a developer, only a user, so forgive me.

Is this a genuine and justified feature I'm requesting? Would it be appropriate to have this raised so that Hermes and keep up with and compete with OpenClaw?

https://github.com/NousResearch/hermes-agent/issues/16923

GitHub
[Feature]: claude-cli backend provider just like openclaw · Issue ...

Problem or Use Case I would like to legally use my Claude subscription with Hermes without risking being banned by Anthropic. Anthropic allows Claude CLI usage through "claude -p" as done...

daring valley
#

you can do this by using /claude-code <prompt>

silver sentinel
silver sentinel
daring valley
#

For /claude-code, you do not pick a Hermes provider/auth method from that Anthropic Native provider page.

That page is for using Claude as the main Hermes model provider. /claude-code is different: it delegates the task to the Claude Code CLI installed on your machine, using Claude Code’s own auth and execution path.

Set up Claude Code itself first:

npm install -g @anthropic-ai/claude-code

Then log into Claude Code:

claude auth login

You can confirm it is working with:

claude auth status

After that, in Hermes, use:

/claude-code your prompt here

So the answer is: no Hermes provider/auth option is needed for /claude-code. The auth lives in Claude Code, and Hermes calls into that local Claude Code setup when you use the command.

We are not planning to add a separate claude-cli backend provider that wraps Claude CLI as a generic model backend. Anthropic’s policy treats third-party harness usage as extra usage, so that is not the direction Hermes is taking. The supported route for Claude Code subscription usage is delegation through /claude-code, not configuring Claude CLI as a Hermes provider.

#

Feel free to reach out with any other questions in a new thread. Closing this as answered.

#

Or we can do them here.

daring valley
silver sentinel
#

I want to use Hermes with my claude max plan that does not violate anthropics terms and conditions.

In the way that openclaw does it, does it charge me extra-usage?:

https://docs.openclaw.ai/gateway/cli-backends#how-it-works

if I use /claude-code within Hermes, does it charge me extra usage?

I want hermes to be the agent and use my max plan as model underneath it so that I save money on a subscription rather then api pricing from anthropic.

I also ideally dont want to be typing /claude-code every time and it would make sense to be able to set my max plan as the default?

sorry if these are dumb questions, im a beginner

#

Thanks again for responding so fast, I appreciate the help.

daring valley
#

The safe way to use Claude Max with Hermes is /claude-code, because that delegates the task to the official Claude Code CLI using Claude Code’s own auth and execution path.

We are not adding Claude CLI as a generic Hermes model backend/provider, and I would not recommend using any third-party harness that turns claude -p into a drop-in model API for another agent. Anthropic’s policy treats third-party harness usage as extra usage, so that is exactly the pattern to avoid if your goal is to stay inside the normal Claude subscription usage path.

So the distinction is:

/claude-code <prompt> = Hermes delegates a task to Claude Code. Claude Code is still the thing using your Claude Max auth.

“Set Claude Max as my default Hermes model through claude-cli/OpenClaw-style backend” = Hermes would be using Claude CLI as a generic underlying model provider. That is not the path Hermes supports, and it is not what I would point you toward for policy-safe subscription use.

For whether /claude-code counts as extra usage: it is using Claude Code through the official Claude Code CLI path, so it should behave like Claude Code usage on your Max plan rather than Anthropic API billing. Hermes is not turning that into an API key or charging path. That said, Anthropic controls the subscription/usage policy and enforcement, so the final authority on exact limits is Anthropic’s own plan terms.

The practical answer is: install and log into Claude Code, then use /claude-code <prompt> from Hermes when you want Claude Code to handle a coding task.

npm install -g @anthropic-ai/claude-code
claude auth login
claude auth status

Then in Hermes:

/claude-code your prompt here

I understand wanting Hermes to be the main agent with Claude Max underneath it by default, but that is the generic-provider pattern we are not supporting here. For Claude Max subscription usage, /claude-code is the supported integration path.

silver sentinel
daring valley
#

The OpenClaw screenshot is their statement about their integration, not an Anthropic policy page we can rely on for Hermes.

The official Anthropic docs do say Claude Code works with Pro/Max plans and that Claude Code usage counts against the same Claude/Claude Code usage limits:

https://support.anthropic.com/en/articles/11145838-using-claude-code-with-your-pro-or-max-plan
https://support.anthropic.com/en/articles/11647753-understanding-usage-and-length-limits

So for Claude Code itself, yes: usage is part of your Claude Pro/Max usage limits, not normal Anthropic API billing.

The part Hermes is not going to do is wrap claude -p as a generic model backend/provider underneath Hermes. That turns the Claude CLI into a third-party harness for another agent, and we are not comfortable treating another project’s screenshot/private statement as enough policy coverage for Hermes.

That is why the supported Hermes path is still:

/claude-code your prompt here

That delegates the task to the official Claude Code CLI using Claude Code’s own auth and execution path. You do not configure it as a Hermes provider, and you do not set Claude Max as the default Hermes model through a claude-cli backend.

So the practical answer is:

Use /claude-code when you want Claude Code to handle the task through your Claude Code login.

Do not expect Hermes to add an OpenClaw-style claude -p model backend.

silver sentinel
#

Ah I see. The screenshot is what OpenClaw said and is not reliable for you to source.

I see it as beneficial if Hermes/Nous-Research team would be able to communicate with Anthropic directly to clarify if "claude -p" would qualify as a breach of policy 3rd party, or if it would be allowed.

Do you think this would be beneficial?

daring valley
#

@astral tangle has spoken with Anthropic and can explain this better than I can.

silver sentinel
silver sentinel
# daring valley <@387972437901312000> has spoken with Anthropic and can explain this better than...

Also, from what i understand, when I use "/claude-code" within Hermes, Hermes thinks and decides what to do and then feeds it into claude code.

This is not what I want. This means I will have to pay for another provider or model for Hermes to do its actual thinking before delegating/feeding/injecting to claude code (which will only then use my max plan)

This is not what I want. I am trying to avoid paying for another provider. (I believe as I have shown, openclaw accommodates my request to not pay for an additional provider beyond my max plan), but I obviously want to use Hermes📈 😁

Let me know if I made any incorrect statements or I have a misunderstanding.

daring valley
silver sentinel
daring valley
#

I'll leave this open for any extra commentary or corrections that @astral tangle may want to add, but I believe I've answered the questions.
Use /claude-code for the -p behavior.
Using this as a circumvention of the 3rd party harness policies is not something we're looking to do.
No, -p will typically not be sent to extra usage unless you try and shove the entire 3rd party harness through it as well. I would expect that they would detect on that and send it to extra usage if we tried, but again, we have no plans to do so. This has been the case since, I believe it was April 4th.

#

Note for Teknium: we may want to make this clear in docs so we stop getting questions like this surrounding confusion on openclaw's implementation.

silver sentinel
daring valley
#

No, Hermes does not currently have an OpenClaw-style configuration that makes Claude Code / claude -p the default backend for the whole agent, and that is not a feature-parity goal for Hermes.

The request is understood. You want Hermes to behave like OpenClaw here: use Claude Code / Claude Max as the underlying default model path so you do not need to pay for another Hermes model provider and do not need to type /claude-code.

That is not how Hermes is designed to use Claude Code.

Hermes has two supported paths:

Use a normal Hermes model provider for the main Hermes agent loop. That can be Nous, OpenRouter, local models, Anthropic Native, etc. If you use Anthropic Native as the main Hermes provider, that is Anthropic provider/API usage and is billed as extra usage; it is not taken from your Claude Pro/Max plan limits.

Use /claude-code <prompt> when you want Hermes to delegate a task to the official Claude Code CLI. That path uses Claude Code’s own auth and execution model.

What Hermes is not adding right now is an OpenClaw-style claude-cli / claude -p provider backend that makes Claude Max the default model underneath Hermes.

If what you want is “Claude Code should be the actual agent and I only want to use my Claude Max plan,” then the cleanest fit is to use Claude Code directly. If what you want is “Hermes should be the agent,” then Hermes needs a normal primary model provider, and /claude-code is available as a delegation tool when you want to hand a task to Claude Code.

OpenClaw may make a different product decision. Hermes is not trying to maintain feature parity with OpenClaw, and OpenClaw’s docs are not the basis for Hermes provider policy.

silver sentinel
#

Before I wanted Claude Max as the default underlying model path so I didn't need to pay for another Hermes model provider. You explained that i not happening. Thank you for that.

Now I have a new request: Are we able to make the Hermes delegate to official Claude Code CLI behavior default using a config file. Openclaw gives an example of this:

https://docs.openclaw.ai/gateway/cli-backends#example-configuration

This way I still pay for Hermes provider. Hermes still uses provider. Hermes delegates to Claude code CLI (which doesn't violate anthropic policy). The only difference is this is default behavior instead of my having to type a skill file /claude-code every single time i send a prompt. This is the only benefit i am requesting.

I think this would provide benefit because if i have autonomous looping/iterative behavior of Hermes and it's subagents, it has the harness & memory benefits of Hermes (using a Hemes provider model which i pay for), but it always uses my Max Plan and Claude Code CLI to perform most of the heavy lifting and execute the actual task. Hermes is just the orchestrator with the harness benefits at that point. This way I don't have to type /claude-code every time, it just happens automatically.

Is this possible? Let me know if I didn't articulate well and I need to provide clarification

#

I'm just looking for a more integrated/automatic/built-in behavior compared to typing/running a skill-file every time

daring valley
#

You articulated it clearly.

What you are asking for now is automatic Claude Code delegation: Hermes would still have a normal primary provider, but every prompt would be treated as if it were sent through /claude-code by default, without you typing /claude-code.

Hermes does not currently support that configuration.

The supported behavior is explicit delegation:

/claude-code <prompt>

That boundary is intentional. Claude Code delegation is available when you ask for it, but Hermes is not currently designed to make Claude Code the default execution path for every prompt or autonomous loop.

So the answer is:

No, there is not currently a Hermes config option equivalent to OpenClaw’s CLI-backend configuration.

No, Hermes is not aiming for OpenClaw feature parity here.

Yes, the current supported path is still /claude-code <prompt> for tasks you want Claude Code to handle.

I understand the benefit you are asking for: Hermes as orchestrator, Claude Code doing most of the task execution automatically. That is a valid product idea to want, but it is not supported today and it is not something I can tell you is planned.

silver sentinel
#

Amazing! Thank you for the support and your time. I appreciate it.

For me particularly it is a difficult decision between the self-inproving features of Hermes and the Claude Code/Claude Agent SDK integration (powered by Claude Max Plan) of OpenClaw.

I wish I had the best of both worlds.

I'll leave this here as documentation, for the record. Unless anyone had anything else to add, I'm good with closing this.