An AI Assistant That Reads Your Terminal — Without Running Anything Behind Your Back

There are two common ways to put an AI assistant in a terminal, and both have a problem.

The first is a chat box that knows nothing about your session. You describe your error, badly, because you are paraphrasing a stack trace you can see and it cannot. It answers generically. This is a web chat window with extra steps.

The second is an agent with a shell. It has full context and can act on it — and it can also run rm against production because it misread a prompt. On a machine you administer, that is a meaningful amount of trust to hand over for a convenience.

Kubexy sits deliberately between the two: full read access to your session, zero autonomous execution.

It reads what you see

Kubexy is a panel alongside your terminal, and it reads the scrollback of the attached session as context. The error you are looking at, the command that produced it, the config you just printed — all of it is available without you retyping or summarising anything.

The scrollback is ANSI-stripped before it is sent, so the model receives the text you can read rather than a soup of colour escape sequences. In practice that is the difference between a useful answer and a confused one, since raw terminal output is largely control codes by volume.

This turns the useless version of the question into the useful one. Instead of "why does my nginx config fail to reload", the question is asked with the actual error, the actual config snippet, and the actual command in view.

It proposes; you run

When Kubexy suggests something executable, it comes back as a command block with a Run button that inserts the command into the attached terminal. Nothing happens until you click.

Being precise about the architecture, because this is the part that matters: Kubexy is tool-less. It has no ability to execute commands, read files, or reach the network on its own. It is not an agent with restricted permissions or a confirmation dialog in front of a capability — it simply does not have the capability. The only path from a suggestion to your shell runs through a click you make.

That is a real constraint and it costs something. Kubexy cannot iterate on its own — run a diagnostic, read the output, adjust, try again — the way an agentic assistant can. What you get in exchange is that no failure mode of the model can result in a command running on your server. For a tool that lives on machines you are responsible for, we think that is the correct trade.

The Kubexer Kubernetes IDE takes the other approach where it makes sense — native tool-calling against live cluster data with mutating actions confirmation-gated — because reading a cluster through the API server is a different risk profile from executing shell commands as root. Both are described in the Kubernetes AI assistant guide.

Bring your own key

Kubexy is BYOK: you supply an API key for a provider you already have an account with. Anthropic, OpenAI, Google, and OpenAI-compatible endpoints are supported, so a self-hosted or proxied model works too.

What this means concretely:

• Your key is stored encrypted on your device through the operating system's secure storage. It is not held on our servers.
• Requests go from your machine to your provider. We are not a middleman for your session contents, and there is no Kubexer-hosted inference layer between you and the model.
• You control the model and the cost. Usage bills to your provider account at your rates, and you choose which model to point at.

The honest caveat: your scrollback is sent to whichever provider you configure, under their terms. That is inherent to using a hosted model, and it is why the choice of provider — including an internal endpoint — is yours rather than ours.

Using it well

Where it earns its place:

• Decoding unfamiliar output. A cryptic error from a subsystem you touch twice a year.
• Reconstructing a command. You know what you want; you do not remember the flags. Faster than the man page and the command is right there to check.
• Explaining someone else's script before you run it.
• Getting unstuck on syntax — awk, jq, find, and the other tools everyone re-learns each time.

Where to be careful: it is a language model reading terminal output, and it can be confidently wrong. The command block is inserted at your prompt precisely so you read it before pressing Enter. Treat a suggestion as a well-informed proposal from someone who can see your screen, not as an authority — particularly for anything destructive.

Kubexy is a Pro feature. Kubexer Terminal is available for macOS, Windows, and Linux.