Skip to content
Axocoatl

Providers

Providers

Section titled “Providers”

Axocoatl wires six LLM providers into the daemon. Each agent picks its own provider and model, so you can route cheap, high-volume steps to a local model and reserve a frontier model for the one step that needs it.

The six providers

Section titled “The six providers”

All six are constructed in setup_providers at daemon bootstrap. A provider is registered only when its config block is present and its key is non-empty.

| provider: value | Backend | |---|---| | ollama | Local Ollama (or any Ollama endpoint) | | openai | OpenAI — or any OpenAI-compatible server (see below) | | anthropic | Anthropic Claude | | gemini | Google Gemini | | mistral | Mistral | | openrouter | OpenRouter (openrouter.ai/api/v1) |

OpenRouter reuses the OpenAI client pointed at openrouter.ai/api/v1 with the provider id openrouter, so you select it with provider: openrouter.

Configuring a provider

Section titled “Configuring a provider”

Each provider block carries an api_key. Use ${ENV_VAR} interpolation — the value is read from the environment (or your .env) at load time, and keys are held as secrets that never appear in debug output.

agents:
  - id: assistant
    name: "Assistant"
    provider: anthropic
    model: claude-sonnet-4-6


providers:
  anthropic:
    api_key: "${ANTHROPIC_API_KEY}"

Ollama is the exception — it takes a base_url, not a key:

providers:
  ollama:
    base_url: "http://localhost:11434"

Per-agent provider + model

Section titled “Per-agent provider + model”

provider: and model: are set on each agent, not globally. A shared provider is registered once; each agent’s configured model is sent as a per-request override, so two agents on the same provider can run different models.

agents:
  - id: triage
    provider: ollama          # cheap, local — high-volume classification
    model: llama3.2
  - id: synthesizer
    provider: anthropic       # frontier — the one hard step
    model: claude-sonnet-4-6


providers:
  ollama:
    base_url: "http://localhost:11434"
  anthropic:
    api_key: "${ANTHROPIC_API_KEY}"

This cheap-local-plus-frontier pattern is the point of per-agent selection: the easy steps cost a fraction of the frontier step. See the multi-provider example below for a runnable, cost-broken-down version.

OpenAI-compatible servers (LM Studio, MLX, vLLM)

Section titled “OpenAI-compatible servers (LM Studio, MLX, vLLM)”

The openai provider honors a configurable base_url. When set, requests go to that endpoint instead of api.openai.com, so you can target any OpenAI-compatible server — LM Studio, MLX/oMLX, vLLM, and similar — while still using provider: openai. The URL must include the API-version suffix the server expects (usually /v1).

agents:
  - id: local
    provider: openai
    model: "your-local-model-name"   # sent per-request to the server


providers:
  openai:
    api_key: "${LOCAL_API_KEY}"      # many local servers accept any non-empty value
    base_url: "http://localhost:1234/v1"   # e.g. LM Studio

Because the agent’s model is sent per request, one OpenAI-compatible provider block can serve several agents each running a different local model.

Rate-limit fallback

Section titled “Rate-limit fallback”

A provider can declare a fallback chain. When the primary provider returns a rate-limit error, the registry retries the request against the next provider in the chain, in order. Non-rate-limit errors are returned immediately (no fallback). Configure it with the provider’s fallback field:

providers:
  openai:
    api_key: "${OPENAI_API_KEY}"
    fallback: "anthropic"
  anthropic:
    api_key: "${ANTHROPIC_API_KEY}"

Local-only

Section titled “Local-only”

Point an agent at ollama with a local base_url and supply no provider keys at all — prompts and data never leave the machine. See the local-only config in the examples gallery.

Verified

Section titled “Verified”

All six providers (Ollama, OpenAI, Anthropic, OpenRouter, Mistral, Gemini) work end-to-end, including the multi-turn tool-calling round-trip, in both chat and resumable sessions.

Section titled “Related”