Setting Ollama

This guide shows you how to use alternative AI models with Claude Code and Quack — from free cloud models via Ollama to paid coding plans from Chinese AI labs. Save your Claude quota for production work, and run experiments on cheaper (or free) models.

Why Use Alternative Models?

Claude is unmatched for complex architecture and production code. But experiments, side projects, and quick iterations burn through your quota fast. Alternative models let you:

• Extend your Claude Max plan to last the full month
• Run experiments for free via Ollama's cloud tier
• Get coding plans from $3/month with surprisingly capable models

Step 1: Install Ollama

Before using alternative models in Quack or Claude Code, you need Ollama running on your machine. Ollama acts as a local proxy that speaks the Anthropic API protocol.

Install

Download from ollama.com, or use your package manager:

macOS:

bash
brew install ollama

Linux:

bash
curl -fsSL https://ollama.com/install.sh | sh

Windows: Download the installer from ollama.com and run it.

Ollama starts as a background service on localhost:11434.

Sign In (Required for Cloud Models)

Cloud models are hosted on Ollama's infrastructure — free, no GPU needed. To access them, create an account on ollama.com, then:

bash
ollama signin

Verify It's Running

bash
curl http://localhost:11434/api/tags

You should see a JSON response. If you get a connection error, start Ollama manually.

Test a Cloud Model

Try running a model to make sure everything works:

bash
ollama run kimi-k2.5:cloud

Cloud models have the :cloud suffix. They run on Ollama's servers, so they don't use your local resources.

Step 2: Configure in Quack

Once Ollama is installed and running, Quack can connect to it with a few clicks.

Open Settings

Click Settings in the bottom-left corner, then navigate to Claude Code.

Select Provider

Under LLM Provider, change the Provider dropdown from Anthropic to Ollama (Local).

The Base URL auto-fills to http://localhost:11434. If Ollama is running, you'll see a green Connected status.

Pick a Model

Choose from the Model dropdown or type a model name manually. Recommended cloud models:

• kimi-k2.5:cloud — Best for TypeScript/React work
• minimax-m2.5:cloud — Fast code reviews and refactoring
• glm-5:cloud — Solid all-rounder

Start a New Chat

The provider switch takes effect on the next chat. You'll see the model name in the status bar at the bottom.

Important: You cannot switch providers mid-session. Always start a fresh chat when changing models.

Alternative: Claude Code CLI

If you prefer the terminal (without Quack), two environment variables are all you need:

bash
ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_AUTH_TOKEN=ollama claude --model kimi-k2.5:cloud

Claude Code connects to Ollama instead of Anthropic. Same interface, same tools — different brain.

Save Yourself Some Typing

Add an alias to your ~/.zshrc (or ~/.bashrc):

bash
alias claude-ollama='ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_AUTH_TOKEN=ollama claude'

Then:

bash
claude-ollama --model minimax-m2.5:cloud

Recommended Models

Kimi K2.5 (Moonshot AI, Beijing)

bash
ollama run kimi-k2.5:cloud

Best for: TypeScript, React components, code pattern understanding. Moments where you forget you're not using Claude.

MiniMax M2.5 (Shanghai)

bash
ollama run minimax-m2.5:cloud

Best for: Fast responses, code reviews, refactoring. Solid reasoning for quick tasks.

GLM-5 (Zhipu AI)

bash
ollama run glm-5:cloud

Best for: General tasks, documentation, analysis. Consistently good at everything.

Going Serious: Paid Coding Plans

When you need guaranteed availability and higher rate limits, each provider offers dedicated subscriptions.

GLM (Zhipu AI / Z.ai) — From $3/month

The cheapest option. Visit z.ai/subscribe:

bash
ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic \
ANTHROPIC_AUTH_TOKEN=your-zhipu-api-key \
claude --model glm-4.7

MiniMax — From $10/month

Visit platform.minimax.io:

bash
ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic \
ANTHROPIC_AUTH_TOKEN=your-minimax-api-key \
claude --model MiniMax-M2.5

Kimi (Moonshot AI) — Pay Per Token

No subscription — pay as you go:

• Input: $0.60 / million tokens
• Output: $2.50 / million tokens

For comparison, Claude Opus charges $5/$25 per million tokens. Kimi is ~10x cheaper.

bash
ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic \
ANTHROPIC_AUTH_TOKEN=your-moonshot-api-key \
claude --model kimi-k2.5

Cost Comparison

Recommended setup: Claude Max for production + MiniMax Starter ($10) for experiments. Total: ~$100/month for practically unlimited AI coding.

Gotchas

1. Install Ollama first. Quack needs Ollama running locally before it can connect. No Ollama = no alternative models.
2. No mid-session switching. Start a fresh chat when changing providers. Mixing providers causes invalid thinking block signature errors.
3. Cloud models need ollama signin. Without it, you only get local models.
4. Non-Claude models may have reduced context, no extended thinking, and variable tool-use quality. Quack shows a warning banner about this.
5. Rate limits on free tier. Ollama's cloud models are rate-limited. For heavy use, get a paid plan.