Skip to main content
Nika’s stdlib v0.1 ships canonical providers — local / self-hosted runners, cloud APIs, and deterministic mock. You select one with a single model: <provider>/<name> field. Local first · nothing leaves your machine unless a cloud provider is explicitly selected.
Canonical source: stdlib/providers-v0.1.md in the spec (counts from canon.yaml). The engine additionally carries an extended catalog of TOML records (crates/nika-catalog/data/llm-providers.toml) used for capability metadata — anything outside the canonical 14 routes through the openai + base_url escape hatch, not its own prefix.

​
Overview

Providers

canonical — local · cloud · mock

API dialects

3 — openai-chat (the 5 local + most cloud) · anthropic · gemini

Capability rules

in model-capabilities.toml

​
How to pick a provider

1

Pick your trust model

Local / self-hosted (Ollama / LM Studio / llama.cpp / LocalAI / vLLM) — zero key, zero egress, sovereign by default. EU cloud (Mistral). US frontier (Anthropic / OpenAI / Google). Aggregator (OpenRouter). Each has different residency and compliance trade-offs.
2

Check the `api_dialect`

All 5 local providers + most cloud providers speak openai-chat — same request shape, differing base URL. anthropic and gemini have native dialects.
3

Verify the backend is reachable

Local providers need their server running (e.g. the Ollama daemon on localhost:11434) — no key. Cloud providers declare an env_var (e.g., MISTRAL_API_KEY). nika doctor checks provider readiness before run.
4

Pick default vs cheap model

Every provider declares both. default_model is the capable tier; cheap_model is the cost-optimized tier. Capability rules in model-capabilities.toml determine which models support vision / tools / streaming / structured output.

​
By category

IDNameDialectEnv varExample model
ollamaOllama daemon (localhost:11434)openai-chat(none)llama3.1
lmstudioLM Studio (localhost:1234/v1)openai-chat(none)qwen2.5-14b-instruct
llamacppllama.cpp llama-server (localhost:8080/v1)openai-chat(none)qwen2.5-7b-instruct
localaiLocalAI (localhost:8080/v1)openai-chat(none)llama-3.1-8b-instruct
vllmvLLM OpenAI server (localhost:8000/v1)openai-chat(none)meta-llama-3.1-70b-instruct
The sovereign default. No API key, no cloud egress — model: ollama/llama3.1 runs offline / air-gapped. Any other OpenAI-compatible local server (Jan · llamafile · KoboldCpp) routes through the openai + base_url escape hatch.

​
Full alphabetical list

anthropic · deepseek · gemini · groq · llamacpp · lmstudio · localai · mistral · mock · ollama · openai · openrouter · vllm · xai total — local · cloud · test · matches canon.yaml in the spec repo. The engine’s extended capability catalog ( TOML records) is an implementation detail — extended IDs are reached via the openai + base_url escape hatch, never their own prefix.

​
Example: invoke a provider

nika: v1
workflow: multi-provider

tasks:
  - id: primary
    on_error: { skip: true }   # a dead provider skips (error preserved) · the gate below reads it
    infer:
      model: ollama/llama3.1                 # local · zero key
      prompt: "Explain linear algebra in one sentence."

  - id: fallback
    depends_on: [primary]
    when: "${{ tasks.primary.error != null }}"   # skip preserves the error · the fallback reads it
    infer:
      model: mistral/mistral-large
      prompt: "Explain linear algebra in one sentence."
infer: is live in v. Run nika doctor first to see which provider keys or local servers are available without printing secrets.

​
See also

Capability rules

The rules that govern which models support which features.

Concepts · Providers

How provider routing works at runtime.

Error codes

NIKA-1xx provider errors.

Schema

Workflow envelope, .nika.yaml top-level keys.