QUESTION - THIS WILL WORK WITH LOCAL AGENTS LIKE QWEN OR GEMINI ?

[xerudro]

This system will work if i will provide LMStudio endpoint and API key, with local models like QWEN 3 30B or Gemini models ? Or with Ornith 6B models, using Hermes as "OS" ?

[zwliJay]

Hi @xerudro, yes, this should be possible.

The important part is how you configure the model credentials for SkillHone.

SkillHone skills can be triggered by any compatible skill-capable runtime, including Hermes, as long as the runtime supports the required skill format and tool-calling workflow.

For the optimization loop, SkillHone currently uses Claude Code / Anthropic-compatible model access under the hood. So for LM Studio, you should point the SkillHone config to LM Studio’s Anthropic-compatible endpoint.

For example, after starting the LM Studio local server, you can configure ~/.skillhone/settings.json roughly like this:

{
  "improver": {
    "api_base": "http://localhost:1234",
    "model": "<your-loaded-lmstudio-model>",
    "max_turns": 100,
    "env": {
      "ANTHROPIC_BASE_URL": "http://localhost:1234",
      "ANTHROPIC_API_KEY": "lmstudio"
    }
  },
  "executor": {
    "api_base": "http://localhost:1234",
    "model": "<your-loaded-lmstudio-model>",
    "workers": 4,
    "max_iterations": 150,
    "context_size": 40000,
    "env": {
      "ANTHROPIC_BASE_URL": "http://localhost:1234",
      "ANTHROPIC_API_KEY": "lmstudio",
      "ANTHROPIC_MODEL": "<your-loaded-lmstudio-model>"
    }
  }
}

So the overall flow is roughly:

Hermes / compatible agent runtime
  -> triggers SkillHone skills
  -> SkillHone optimization loop uses the configured improver model
  -> Claude Code talks to LM Studio through the Anthropic-compatible endpoint
  -> updated SkillHone skills
  -> usable again by Hermes / compatible runtimes

Qwen 3 30B should be a reasonable local model to try. Smaller models such as Ornith 6B may work for simple cases, but the optimization loop usually benefits from stronger reasoning, coding, and tool-use ability.

Gemini is not local in the LM Studio sense, but it can also work if configured through an Anthropic-compatible provider/API.

So in short: yes, Hermes can be used as the runtime/OS side for triggering the skills, and LM Studio/local models can be used by putting the Anthropic-compatible endpoint and model name into ~/.skillhone/settings.json.