@arnilo/prism-provider-ollama

Ollama Cloud / local provider package for Prism, over the OpenAI-compatible POST {base}/chat/completions route.

import { createOllamaProviderPackage, listOllamaModels } from "@arnilo/prism-provider-ollama";

// Caller-gated discovery (never during setup) — no hard-coded model catalog.
const models = await listOllamaModels({ apiKey: "fake-ollama-key", preset: "cloud" });

api.registerProviderPackage(createOllamaProviderPackage({
  apiKey: "fake-ollama-key", // omit for unauthenticated local `ollama serve`
  preset: "cloud", // or "local" (http://localhost:11434/v1) or explicit baseUrl
  models,
}));

Exports:

• createOllamaProviderPackage() / createOllamaProvider()
• listOllamaModels() / mapOllamaModel() / defineOllamaModel()
• ollamaBaseUrl() / DEFAULT_OLLAMA_BASE_URL / OllamaBasePreset
• ollamaBody() / ollamaEvents() / ollamaReasoningEffort()

Behavior:

• Dynamic model discovery via OpenAI-compatible GET {base}/models (native GET {base}/api/tags documented as alternate); nothing hard-coded.
• Cloud (https://ollama.com/v1, Bearer API key) and local (http://localhost:11434/v1, typically unauthenticated) base-URL presets.
• reasoning_effort passthrough (e.g. gpt-oss).
• Implicit cache only: Ollama reuses its KV/prompt cache automatically with no request knob and no cached-token count, so Usage.cacheReadTokens is intentionally left undefined (documented ceiling).

Security defaults:

• No network calls during import, setup, build, or default tests.
• No automatic environment, file, keychain, or shell credential lookup.
• Cloud API key sent only as Authorization: Bearer; redacted from all errors.
• No local filesystem paths in request payloads.

See docs/providers/ollama.md for the full API page.