Ollama

Ollama runs local models on your machine. agentfootprint's ollama() factory wraps Ollama's OpenAI-compatible endpoint as an LLMProvider — your agent code is identical between local-Ollama dev and cloud-Anthropic production.

Install

Ollama itself: see ollama.com/download. Pull a model:

ollama pull llama3.1

agentfootprint side: ollama() uses the openai SDK under the hood (via OpenAI-compatible endpoint), so install the OpenAI peer dep:

npm install openai

Use

import { Agent } from 'agentfootprint';
import { ollama } from 'agentfootprint/llm-providers';

const provider = ollama({
  defaultModel: 'llama3.1',         // used when the request model is the 'ollama' shorthand
  host: 'http://localhost:11434',   // default — talks to http://localhost:11434/v1
});

const agent = Agent.create({
  provider,
  model: 'llama3.1',
}).build();

ollama() is a thin convenience over openai() — it lives on the agentfootprint/llm-providers subpath (vendor-SDK-backed providers are kept off the main barrel for tree-shaking). Options: host (defaults to http://localhost:11434), baseURL (overrides host), apiKey (Ollama ignores it; defaults to 'ollama'), defaultModel (defaults to llama3.2), and defaultMaxTokens.

When to use Ollama

• Local development — no API key, no cost, no network.
• Privacy-sensitive workloads — model + data never leave the machine.
• Cost-sensitive batch jobs — local inference for high-volume background processing.
• Edge / offline deployments — bundle the model with your app.

Limitations

• Local models are smaller and less capable than frontier models. Tool-calling reliability varies by model — llama3.1 + qwen2.5 + deepseek work; smaller models may struggle.
• No streaming for some local models (depends on Ollama config).
• Multi-modal not exposed.

Switching to Anthropic / OpenAI for production

Same agent code; swap the provider one line:

import { anthropic, ollama } from 'agentfootprint/llm-providers';

const provider = process.env.NODE_ENV === 'production'
  ? anthropic({ apiKey: process.env.ANTHROPIC_API_KEY! })
  : ollama({ defaultModel: 'llama3.1' });

Next steps

• OpenAI — baseURL pattern works with any OpenAI-compatible endpoint
• Mocks-first development — mock() for $0 deterministic tests