AI, the comic

Advisors, rules and a back door.

The same kitchen as the other comics, now with more than one advisor. One from next door, one who reads the recipe of a cake that collapsed, a board of rules for when they may lend a hand, and a visitor at the back door who has to show his ticket.

1 An advisor from next door

Claude is not the only advisor. Take one from your own street: a model on your own machine, such as Ollama, Docker Model Runner or LM Studio. No bill per question, and nothing has to leave the house.

Before you rely on him, a short test: one question that needs a tool. The app says how fast he answered, and whether he can use tools at all. A small advisor is quick, but often beside the point.

An advisor from out of town — an address beyond this computer and your own network — needs https and your yes in a window of the app first. A key you give him goes only to that one address.

2 When the cake collapses

A build that fails is a cake that collapsed in the oven. The app keeps what went wrong: the recipe as it is now (the Dockerfile), the last lines from the oven, and what BuildKit said about the recipe.

Copy that as a diagnosis for an issue, with the secret ingredients left out, or ask the advisor. He reads the recipe and the oven’s last words, and points at the step that broke.

He says what to change, with the lines as they should be. He changes nothing and bakes nothing himself: the next cake is yours.

3 A pair of hands, if you want

Out of the box an advisor only suggests. In the settings you can let him act: strict, where every action waits for your yes, or balanced, where what your rules allow happens right away and the rest still asks.

Some things are never his job. A container with the label docker-client-mx.ai=never, or a machine you named, stays out of reach. Removing, prune, volumes, files and typing in the terminal do not exist for him at all.

Whatever he does becomes a notification. What reaches your phone says only what happened and where, never words from the model.

4 A visitor at the back door

Other AI tools, such as Claude Code, can come to the back door too: the MCP endpoint. They may look around: the machines, the containers, the logs and the measurements, with the secrets left out.

Only with the token, and only from this house: the door is on 127.0.0.1, and a browser from the street is turned away. The token stays in the app; you copy it to the visitor once.

If he wants to act, the same rules apply, only on the machine the app is on, and the app asks you in a window first, every time — unless you turn that off yourself.

In Docker Client MX: an advisor from Anthropic or from next door, help when the cake collapses, a pair of hands only within your rules, and a back door with a lock.

The cheat sheet

Every picture above, in the app’s words.

The advisor from next door: A local model: Ollama, Docker Model Runner, LM Studio
Passing the cup: The test: how fast it answers, and whether it can use tools
The stranger from out of town: An address beyond your network: https and your yes first
The cake that collapsed: A failed build, with a diagnosis or a conversation about it
The rules on the board: What the AI may do: suggest only, strict or balanced
The padlock on Ada’s counter: The label docker-client-mx.ai=never, and machines the AI never acts on
The buzz in your pocket: A notification for everything the AI did
The visitor at the back door: An MCP client such as Claude Code
The ticket: The token of the MCP endpoint, only on 127.0.0.1

Bring your own advisor.

Claude or a model on your own machine, help where something fails, hands only within your rules, and an MCP endpoint with a lock.

What the AI does in the app →Claude, the comic →Download for Windows