Skip to main content
The language names canonical providers: local servers, cloud APIs and deterministic mock. The binary drives exactly those. You select one with a single model: <provider>/<name> field. Your choice: a local provider keeps inference on your machine; a cloud provider runs only when a workflow names it.
Canonical source: stdlib/providers-v0.1.md in the specification the binary embeds. The engine’s catalog (llm-providers.toml at v0.123.0) holds records: nika catalog lists the it cannot drive under CATALOG ONLY, and nika check refuses them. Every table below matches nika catalog --json on v.

Overview

Providers

canonical: local · cloud · mock

Wires

openai-chat for most · Ollama native · anthropic · gemini · mock

Capability rules

in model-capabilities.toml

How to pick a provider

1

Pick where inference runs

A local server (Ollama, LM Studio, llama.cpp, LocalAI, vLLM, or Nika’s own nika model serve) needs no key and sends nothing off the machine. A cloud API runs on your key, with that vendor’s residency and compliance terms: Mistral and Scaleway are EU providers; OpenRouter and Hugging Face route many vendors’ models under one key.
2

See what this machine has

nika doctor reports each provider key as set or unset, never its value. nika doctor --ping also connects to the configured local servers and reads their model lists.
3

Pick the model

Each provider declares a default and a cheap model (below). nika catalog --json lists every model with its context window, output cap and capabilities; the capability rules decide its token-limit parameter, reasoning, modalities and JSON mode.
4

Check before you run

nika check prints a MODELS line (the model resolves in this build) and an ACCESS line (this machine has a path to it), without calling the model.

By category

Local first, then cloud in the canon’s presentation order.
No key. NIKA_<ID>_BASE_URL points one at another host or port (host, host:port or a full URL; a bare host keeps the default port), and Ollama also honors OLLAMA_HOST. llama.cpp and LocalAI share port 8080 by default: running both needs one override.
Built-in local inference. nika model serve starts an OpenAI-compatible server on 127.0.0.1:8712 (Qwen3-family models in 0.124.3). A workflow reaches it as llamacpp/<model id> with NIKA_LLAMACPP_BASE_URL=http://127.0.0.1:8712. See Local models.

Catalog only

ai21 · azure · bedrock · cerebras · cloudflare · cohere · databricks · deepinfra · fireworks · hyperbolic · minimax · native · perplexity · qwen · replicate · sambanova · together · vertex · voyage · writer · zhipuThey are records, not routes: nika check refuses a model under any of them, and nika catalog prints them under CATALOG ONLY. Reach an OpenAI-compatible one through NIKA_OPENAI_BASE_URL as above.

Example: a local model with a cloud fallback

A task names one model: there is no fallback: field. The second task runs only when the first one failed, and --max-cost-usd bounds the metered spend.

See also

Capability rules

The rules that decide what each model accepts.

Concepts · Providers

Model and access, the wires, structured output.

Local models

Run every step on your own machine.

Error codes

NIKA-INFER-* for model calls, NIKA-1800 for no ready access path.