Ollama

Ollama is a local model runtime that serves models over HTTP. switchboard-ai-sdk discovers the Ollama server, lists installed models, and normalizes chat and health checks.

How discovery works

import { discover } from "switchboard-ai-sdk";

const tools = await discover();
const ollama = tools.find((t) => t.id === "ollama");

console.log({
  available: ollama?.available,
  models: ollama?.models,
  defaultModel: ollama?.defaultModel,
});

Discovery calls the Ollama HTTP API at the configured host (default http://127.0.0.1:11434) and returns all installed models.

How auth works

Ollama does not require authentication. checkAuth() and startAuth() are not available on Ollama connected tools.

How model selection works

Ollama returns the full list of installed local models. The default model is chosen in this order:

1. Configured ollamaModel or SWITCHBOARD_OLLAMA_MODEL
2. First model returned by Ollama

import { configure } from "switchboard-ai-sdk";

configure({
  ollamaHost: "http://192.168.1.20:11434",
  ollamaModel: "qwen3:14b"
});

You can override per chat:

const result = await tool.chat({
  messages: [{ role: "user", content: "Hello" }],
  model: "llama3.1:latest"
});

Minimal example

import { connect } from "switchboard-ai-sdk";

const tool = await connect("ollama");

if (!(await tool.health())) {
  throw new Error("Ollama is not running.");
}

const result = await tool.chat({
  messages: [
    { role: "user", content: "Explain dependency injection in simple terms." }
  ],
});

console.log(result.message.content);

Common errors

Error	Cause	Fix
ToolNotFoundError	Ollama server is not reachable	Start Ollama or check ollamaHost
ToolUnavailableError	Ollama running but model missing	Pull the model with ollama pull <model>
ProviderExecutionError	Ollama returned an error	Check Ollama logs and retry
TimeoutError	Model loading or generation took too long	Increase timeoutMs or use a smaller model

Limitations

• Ollama must be installed and running separately
• Only models already pulled on the user's machine are available
• The first discovered model may not be the best one; configure a default when possible
• Ollama runs as a runtime, not an agent, so it will not autonomously modify files