Michael Limberger

The CLI that used to phone home

This talk points Claude Code at local Ollama models. You keep the same agentic coding CLI, but the weights live on your LAN.

Ollama added Anthropic API compatibility in version 0.14.0 (January 2026). That means we can point Claude Code at our own local models instead of Anthropic's cloud. Tonight we set that up. You get the polish of Anthropic's CLI without the cloud dependency.

What "vibe coding" actually is

Vibe coding is when you describe what you want and working code appears. Instead of typing every character, you tell the AI your intent: "Create a Perl script that fetches RSS feeds and extracts titles." It writes the script, edits files directly, runs tests, and iterates until it works.

This is not autocomplete. This is an agent that reads your existing code, understands the project structure, makes multi-file changes, executes shell commands, and learns your preferences over time. Claude Code is currently one of the best tools for that workflow. And now we can run it without the cloud.

Two boxes, or one

Here is what we are building. Your machine runs Claude Code, the CLI client. Over the LAN it talks to an AI server running Ollama with qwen3-coder. Or everything lives on one machine, with Claude Code talking to Ollama at localhost:11434.

+------------------+         +------------------+
|  Your Machine    |         |  AI Server       |
|                  |   LAN   |                  |
|  Claude Code     | ------> |  Ollama          |
|  (CLI client)    |         |  (qwen3-coder)   |
+------------------+         +------------------+

Same machine:

+------------------------------------------+
|  Your Machine                            |
|                                          |
|  Claude Code  ------>  Ollama            |
|  (localhost:11434)                       |
+------------------------------------------+

Both layouts work. We will cover a remote Ollama server, since that is the more complex case. Localhost is the same idea with a simpler address.

What we cover while it installs

Installing Ollama and Claude Code. Configuring the connection. How context windows work, and why they matter. Timeouts for local models. Using CLAUDE.md to teach Claude about your project. Power-user features: slash commands and context management. Troubleshooting when it hangs or phones home by mistake.

What you need on the desk

A Mac with Apple Silicon (M1, M2, M3, or M4). 16GB unified memory minimum, 32GB+ recommended. About 30GB free disk. Or a Linux machine with a decent GPU and the same memory and storage story. macOS or Linux (Windows is untested by us). Node.js for the Claude Code installer. Terminal comfort, because we live in the command line.

What this is not

This is a casual walkthrough of something cool we figured out, not a professional workshop with certificates and polished slides. We cover the common cases. You will figure out the edges.

A note on speed

Local inference is slower than cloud APIs. Way slower. Anthropic's cloud often lands around 20 to 50 tokens per second. Local qwen3-coder on an M4 is more like 5 to 15 tokens per second. Expect 30 to 120 seconds for responses depending on complexity. That is the trade for privacy and zero per-token cost.

If speed is your priority, pay for the cloud API. If privacy and independence matter more, welcome aboard. Next: prerequisites.