Available Models¶

The model field in an agent's frontmatter is written as <provider>/<model-id>, where the provider prefix is one of anthropic, mistral, openai, or ollama.

Example:

---
model: anthropic/claude-sonnet-4-6
---

The tables below list the models Knosh can use. Only models with tool-use support are shown — models intended solely for embeddings, moderation, or audio are excluded, as Knosh relies on tool calls to do its work.

Anthropic¶

Requires anthropicApiKey in knosh.json, or the ANTHROPIC_API_KEY environment variable.

Mistral¶

Requires mistralApiKey in knosh.json, or the MISTRAL_API_KEY environment variable.

OpenAI¶

Requires openaiApiKey in knosh.json, or the OPENAI_API_KEY environment variable.

Ollama¶

Ollama lets you run open-weight models locally (or on any reachable host). No API key is required, but you must set ollamaURL in knosh.json to point at your Ollama server.

{
  "ollamaURL": "http://localhost:11434"
}

Use localhost when Ollama is running on the same machine as Knosh. For a remote server, supply the appropriate base URL, for example http://192.168.1.10:11434 or https://ollama.example.com.

The model ID is whatever tag you have pulled in Ollama (e.g. llama3.2, qwen2.5-coder:14b). Write it as:

---
model: ollama/llama3.2
---

Because Ollama models are not in Koog's built-in registry, Knosh constructs a minimal model descriptor for them at runtime with tool support enabled. Any model that supports function/tool calling in Ollama should work.