Local

The local pseudo-provider points Stella at any OpenAI-compatible endpoint you run yourself — Ollama, vLLM, LM Studio, or a llama.cpp server. No API key, no cloud, no per-token bill; your tokens never leave the machine. It is never auto-detected: local inference is always an explicit choice.

Setup

There is nothing to export — no key is required. A bearer token is always sent regardless, because OpenAI-compatible clients must send something: the placeholder local when LOCAL_API_KEY is unset, your token when it is set. Servers that ignore auth (the common case) never notice; servers that enforce it just need LOCAL_API_KEY exported. What is required is the endpoint: --base-url (or its env form STELLA_BASE_URL) must accompany --model local/…, every time.

stella --model local/llama3.3 --base-url http://localhost:11434/v1 chat

That is the Ollama default port; vLLM conventionally serves on :8000/v1, LM Studio on :1234/v1, llama.cpp's server on whatever you gave --port — any of them work, as long as the endpoint speaks the OpenAI-compatible dialect. The model slug after local/ must be a model the server actually has loaded (for Ollama, a pulled model name).

export STELLA_BASE_URL=http://localhost:11434/v1
stella --model local/llama3.3 chat        # no --base-url needed now

Skip the retyping: a custom provider id

The id local is reserved — it cannot be redefined in settings.json, which is why --base-url is always required with it. To bake the endpoint into config instead, define your own provider under a non-reserved id with a base_url; the dialect defaults to openai-compatible, so one field is usually enough:

{
  "providers": {
    "ollama": {
      "base_url": "http://localhost:11434/v1"
    }
  }
}

stella --model ollama/llama3.3 chat

See settings.json for the full providers schema — custom providers get the same treatment as built-ins, including a default_model. Two more knobs worth knowing:

• Credentials. A config-defined provider reads its key from a derived env var: the id uppercased, with anything outside [A-Za-z0-9] folded to _, plus _API_KEY — my-gateway reads MY_GATEWAY_API_KEY, the ollama entry above reads OLLAMA_API_KEY. Set api_key_env on the entry to name a different var.
• Dialect. dialect may be openai-compatible (the default), openai-responses, anthropic, or gemini — so a custom id can front an Anthropic-shaped or Gemini-shaped gateway too. vertex and bedrock are rejected for custom ids: those dialects need credential resolution (project/region addressing, SigV4 signing) that a settings entry cannot express.

In the pipeline

Local models carry zero dollar cost, which rewrites the role calculus: triage on a local model is free classification, and a local worker is unlimited iteration on tasks it can handle. Quality depends entirely on the model you serve — a 70B coder is a credible budget worker; a 7B model is a triage-and-drafts tool. Judge duty is where local models most often fall short; keep a frontier judge from the Model guide over any local worker. Mix local and cloud per agent with agent_engine_config, and see Examples for hybrid profiles.

Pricing

Catalog prices do not apply — there are none. Runs on local/… record $0 in the Observatory; your costs are hardware and electricity. Token and latency telemetry are still recorded in full, so you can compare a local model's resolve rate head-to-head against paid ones.