Skip to content
Hooman

Models

Hooman is bring-your-own-model by design: local llama.cpp / MLX with no API keys, hosted providers with your keys, or OpenAI-compatible baseUrls pointed at private gateways and vLLM clusters. First-run setup (CLI and VS Code) validates one provider and writes its available chat LLMs into config.json. That file splits model configuration into reusable providers (credentials and provider-level options) and llms (named model presets that reference a provider by name). Each provider page below covers its options fields, reasoning support, and example configs.

Supported providers

Section titled “Supported providers”
Provider Runtime id
Anthropic anthropic
Azure azure
Bedrock bedrock
Google google
Groq groq
llama.cpp llama-cpp
MiniMax minimax
MLX mlx
Moonshot moonshot
Ollama ollama
OpenAI openai
OpenRouter openrouter
xAI xai

Shared shape

Section titled “Shared shape”
{
  "name": "MiniMax",
  "provider": "minimax",
  "options": {
    "apiKey": "..."
  }
}
{
  "name": "MiniMax M3",
  "provider": "MiniMax",
  "options": {
    "model": "MiniMax-M3",
    "temperature": 1,
    "maxTokens": 4096
  },
  "default": true
}

Every llms entry, regardless of provider, carries the same normalized options shape: model (required), optional temperature, optional topP, optional maxTokens (Google maps this to the SDK’s maxOutputTokens), and optional context (context size in tokens — only honored by the local llama.cpp and MLX providers, where it overrides the provider-level context and feeds the context-usage gauge; other providers ignore it).

All reasoning-capable providers additionally share a common optional reasoning object on the provider options: { effort?, summary?, display? }. effort ("minimal" | "low" | "medium" | "high") enables thinking; summary (OpenAI/Azure Responses API only) controls summary verbosity; display (Bedrock Claude / MiniMax only) controls whether the reasoning trace is returned. Each provider page documents its exact mapping.

Reasoning options

Section titled “Reasoning options”

Provider-level options.reasoning uses the same normalized shape everywhere, even though each backend maps it differently:

Field Type Used by
effort `“minimal” “low”
summary `“auto” “concise”
display `“summarized” “omitted”`

Notes:

  • reasoning belongs on the provider entry, not the per-LLM options block.
  • top-level reasoning in config.json is separate: it controls how reasoning is displayed in the UI, not whether the model reasons.
  • Providers that do not support a given field simply ignore it or reject it as documented on their provider page.

LLM metadata

Section titled “LLM metadata”

Each LLM entry may carry an optional metadata block used to display context-window utilization and cumulative session cost in the chat status bar, the VS Code extension footer, and via ACP usage_update:

{
  "name": "Haiku 4.5",
  "provider": "LiteLLM Anthropic",
  "metadata": {
    "name": "claude-haiku-4.5",
    "context": 200000,
    "costs": { "input/m": 1, "cache/m": 0.1, "output/m": 5 },
    "modality": { "text": true, "image": true, "pdf": true }
  },
  "options": { "model": "claude-haiku-4.5" },
  "default": true
}
  • metadata.name is required when the block is present, and is the identifier looked up in the models.dev catalog (cached under ~/.hooman/cache/, refreshed at most once daily). When metadata is omitted, options.model is used as the lookup name.
  • metadata.context (window size in tokens) and metadata.costs (USD per million tokens; "cache/m" prices cached-input reads) override whatever models.dev resolves.
  • metadata.modality can explicitly override the model’s advertised input modalities (text, image, pdf, audio, video). At runtime, attachments, ACP image/audio/blob blocks, and filesystem read_file / read_multiple_files with binary: true all use the resolved modality (config override → models.dev → text-only). Unsupported modalities become diagnostics/text (or base64 for unknown binaries) instead of forcing media blocks the model cannot accept. Editable from /config and the VS Code settings editor.
  • If neither the config nor models.dev yields the data, context usage and cost are simply not shown.
  • For local providers (llama.cpp, MLX, Ollama), catalog costs are never applied — local inference is free, and the catalog prices the hosted API serving the same model id — so only the context window resolves and no $ cost is displayed. An explicit metadata.costs block still wins if you set one.

Resolution merges config-provided metadata over the models.dev catalog: model ids are matched with separator normalization plus boundary containment (so claude-haiku-4.5, anthropic/claude-sonnet-4-5, and Bedrock’s region-prefixed ids all resolve), preferring the provider whose id equals the model’s canonical lab. Per-request cost is computed from additive-shape usage and accumulated at the rates of the model that served each request; once a request with usage runs unpriced, cost reporting stops for the session rather than under-reporting.

Context usage is a property of the conversation, not the model: switching models carries the used tokens over and rescales them against the new model’s window, while accrued cost is kept.