LLM Providers

Providers are how Hoziron connects to the language models that power your agents. The platform is model-agnostic — you can use cloud APIs (OpenAI, Anthropic, Azure), self-hosted models (Ollama, vLLM), or a mix of both. Complexity-based routing lets you send simple tasks to fast, cheap models and reserve expensive ones for work that needs them.

This design means you're never locked into a single vendor. Swap providers, add fallbacks, or migrate to local models without changing your agents or competencies.

What you'll find here

GuideDescription
Configuring providersRegister providers in config.toml, set API keys, verify connectivity
Local modelsSet up Ollama or vLLM, handle Docker networking, manage model pulls
Complexity routingDefine thresholds for automatic model selection based on task complexity

Common tasks

How routing works

Requests classify into one of three tiers (local/economy/standard) purely by size, then resolve against that tier's ranked provider preference — filtered by live circuit-breaker availability. The carrier's PII policy can independently force a request into the local tier's resolution path regardless of size. See Complexity routing for the full mechanism.

Key configuration

The provider/model inventory is a closed, boot-time-only set declared in config.toml — there is no runtime API to add a provider (ADR-053):

[[provider]]
id     = "big.licence"
driver = "anthropic"            # closed set of kernel-shipped drivers
api_key_env = "ANTHROPIC_API_KEY"

  [[provider.model]]
  id = "claude-sonnet-4-5-20260929"

[[provider]]
id     = "ollama.west"
driver = "ollama"                # AlwaysLocal — structurally in-boundary
url    = "http://localhost:11434/v1"

  [[provider.model]]
  id = "llama3.2"

[routing_gateway.standard]
preference = [{ provider = "big.licence", model = "claude-sonnet-4-5-20260929" }]

API keys are read from environment variables at request time — they're never stored in the config file itself. See Configuring providers for the full schema and the closed driver list.

  • Agents — each agent is assigned to (or routed between) providers
  • Deployment § Docker Compose — networking between Hoziron and local model servers
  • Security — protecting API keys and provider credentials