Claude HUD: See Inside Claude Code in Real Time

Agentic coding tools have an observability problem. The model is doing things — reading files, running commands, making decisions about what to do next — and you're watching a stream of text that summarizes some of it. If something goes wrong, you're often debugging after the fact.

Claude HUD is a project that tries to fix this. It's a heads-up display that surfaces what's actually happening inside a Claude Code session in real time: tool calls, file reads, token usage, decision points.

Why this matters

Trust in agentic tools is fragile. The moment a model does something unexpected — deletes a file it shouldn't have, misreads the codebase structure, goes down a wrong path for twenty turns — you lose confidence in the whole system. And if you can't see what it was doing, you can't even diagnose what went wrong.

A HUD that makes the internal state visible addresses this directly. Not because it prevents mistakes, but because it gives you enough information to catch them early and understand them when they happen.

What I use it for

Primarily debugging. When a Claude Code session produces a result I don't expect, the HUD trace tells me where the reasoning went sideways. It's also useful for optimizing — seeing token usage per tool call shows you quickly which operations are expensive and whether there's a cheaper way to express the same intent.

The right mental model is something like browser devtools for your agentic session. You don't run with devtools open all day, but you're glad they exist when you need them.