Hm.
GitHub Get Hm
Docs / Agent

Providers and models

Bring your own key for Anthropic, Gemini, OpenAI, Azure OpenAI or any OpenAI-compatible endpoint.

Hm talks to model providers directly with your own key. Nothing leaves your machine except the calls you configure.

Supported providers#

Type Notes
anthropic Claude models, with an API key
google Gemini, with an API key or on Vertex AI with your gcloud sign-in. Thinking streams, and thought signatures go back with the history as the API requires.
openai OpenAI, or any OpenAI-compatible base_url
azure Azure OpenAI: a deployment with api_version, or the v1 API
openrouter, groq, cerebras, mistral, xai, deepseek, together, fireworks, huggingface, ollama, lmstudio Presets: the base_url is filled in, so Settings → Agent → Add a provider only asks for a key 
agent:
  default_provider: anthropic
  providers:
    openrouter: { type: openrouter, api_key: "${OPENROUTER_API_KEY}" }
    ollama: { type: ollama, model: qwen3, context_window: 32768 }
    gemini: { type: google, api_key: "keychain:gemini", model: gemini-2.5-pro }
    vertex: { type: google, project: "${GOOGLE_CLOUD_PROJECT}", location: us-central1, model: gemini-2.5-pro }
    azure: { type: azure, base_url: "https://my-resource.openai.azure.com", api_key: "keychain:azure", api_version: "2024-10-21", model: my-gpt-5 }
    work: { type: openai, base_url: "https://llm.example.com/v1", api_key: "keychain:work" }

For Vertex AI, sign in once with gcloud auth application-default login. For Azure, model is the deployment's name.

Switching models#

  • ctrl+l, or the composer's model menu → All models…, searches every provider's models.
  • /model <name> picks one by provider/model, model id or part of one.
  • ctrl+p steps through agent.favorite_models.

The history goes along when you switch; thinking written by another provider is left out. A reopened session goes on with the model and thinking it last ran.

Thinking#

/think off|low|medium|high|xhigh|max sets how hard the model thinks for the session, and /think default goes back to the provider's. On OpenAI and Azure it sets reasoning_effort on the reasoning models (o-series, gpt-5).

The model registry#

models.json, beside config.yaml in ~/.config/hm or .hm, adds models or changes what hm knows of them: context window, max output, prices, what they support and their effort levels. The workspace's file wins. The context meter, compaction and the cost of each reply use it, and changes apply as soon as you save.

[
  { "id": "qwen3", "provider": "ollama", "name": "Qwen 3", "context_window": 32768, "max_output": 8192,
    "supports": { "tools": true, "images": false, "thinking": true } },
  { "id": "claude-opus-5-5", "price": { "input": 4, "output": 20, "cache_read": 0.2, "cache_write": 5 },
    "efforts": ["low", "medium", "high", "xhigh", "max"] }
]

Every field but id is optional; one you leave out comes from hm's own table or a shorter id that matches (claude-opus-5 covers claude-opus-5-20261101). provider limits an entry to one name under agent.providers. Prices are USD per million tokens and fill in the cost when the API does not report one.

Subscriptions#

Hm uses API keys and cloud credentials, not consumer subscriptions: most providers' terms don't allow a third-party tool to use a subscription. The notes in the repository list what each provider allows.

Edit this page on GitHub
esc