Session timeline
Four sessions between 12:39 PM and midnight, a tick for every prompt, tokens and cost at the end of each bar.

Every Claude Code session you ran today, on one dark screen: tokens, dollars, diffs, commits and pull requests. An anchor reads it out loud, and you can talk back.
A theme switch by voice, a spoken briefing, a commit diff, and a permission prompt approved out loud.
Claude Code writes a transcript for every session. The HUD parses those files, walks each folder with git, and listens on localhost for live hook events. It never talks to Claude.
~/.claude/projects/**/*.jsonllog, show, diff for every folder touchedoptional token · PR state and sizerelay.js → 127.0.0.1:47821frameless · six themes · any daydigest at 18:00 · finished · needs youQuest browser over your LANnpm run scanThe first screen answers how much and where. Every panel drills down.
A reactor ring for the day's total, split into input, output and cache, with a cost bar per model.
Sessions, prompts, calls, files, projects, commits and PRs, with the day's active span.
Daily tokens and cost. Click a bar to jump to that day; a dial shows the peak hour.
Every session across the day, with a tick for each prompt.
Sessions, commits, working tree and Claude's own edits. Click a commit for the full diff.
PRs from your transcripts. Add a GitHub token for state, size and the ones you opened.


The 2,400 pixels under the first screen answer when, where, what and why.
Four sessions between 12:39 PM and midnight, a tick for every prompt, tokens and cost at the end of each bar.

Every folder Claude worked in, sized by tokens, with commits and uncommitted changes.

Every prompt with its time, the session's vital signs, and each shell command Claude ran.

Rebuilt from Claude's own Edit and Write calls, so it shows work that never got committed.

Seven commits in 48 minutes, and a merged PR linked back to the sessions that wrote it.


$40.74 of the day's $93.58 was avoidable. Every finding names the sessions behind it and ends with the fix.

Each day gets a score out of 100. Each finding comes with a dollar figure and the command that fixes it.
| Finding | Fires when | Fix |
|---|---|---|
| highLong-context sessions | Above 150K context, every turn re-reads it | /compact or /clear |
| highCache rebuilds after breaks | A big session resumes after the cache expired | /compact before a break |
| medBulky tool output | Large results stay in context for every later turn | | tail -50, a subagent |
| medQuick tasks on a premium model | Short sessions on Opus or Fable | /model sonnet |
| highDaily budget | Past 50%, 80% and 100% of your cap | Wrap up, move to Sonnet |
Past ~150–200K context, /compact keeps the decisions and drops the noise.
Recognized offline by Windows, nothing recorded. It only acts on phrases it knows, built from your Start menu and the folders Claude has worked in.
Anchor, brief me
Anchor, show diffs
Anchor, next day
Anchor, theme name
Anchor, open Spotify
Anchor, launch Discord
Anchor, start Claude in orbit ui
Anchor, approve
Anchor, deny



When Claude Code wants to run a tool, the prompt lands in the HUD with a countdown. Click, or say it. If nobody answers, the terminal asks as usual.
PermissionRequest · Bash: npm run buildPOST 127.0.0.1:47821/hookapproval strip · toast · anchor asks“Anchor, approve” or click# relay.js stdout → Claude Code (try the buttons)
waiting for an answer…

One click installs the hooks. Every session then reports to the HUD, from a terminal, VS Code or the desktop app, and a closed HUD never slows one down.
| Hook event | In the HUD |
|---|---|
SessionStart | Session appears in Live sessions |
UserPromptSubmit | Working, with the prompt in the feed |
Notification | Waiting · toast “needs you” |
PermissionRequest | Approval with a countdown |
Stop | Toast “Claude finished” for turns over 20 s |
PreCompact | Compaction logged |
Live recordings of each. Switch with M, from the menu, or by voice.
Quiet, data first. The default.
Green phosphor and beveled chrome, 1999.
Cyan reactor rings and spinning dials.
Ink outlines, hazard tape, Japanese labels.
Ultraviolet print with glowing outlines.
White stock, hot pink and electric blue overprint.
The real HUD renders on your PC and streams to a curved screen in a Meta Quest, over your own network.
It reads transcripts already on disk.
Speech runs on Windows itself.
GitHub, only if you add a token.
A 12 fps clock, no endless animation.
No. It reads files Claude Code already wrote and runs git locally. The one optional network call is to GitHub, if you add a token.
It's what the same usage would cost at API prices. On a flat plan it's still the clearest way to compare sessions and spot waste.
No. Windows' built-in recognizer listens for a fixed list of phrases. Nothing is saved or sent, and voice is off until you turn it on.
The relay can't connect, exits with no output, and Claude Code shows its normal prompt. Same if nobody answers in time.
Not today. Voice and app launching use Windows APIs. The digest itself is plain Node and runs anywhere with npm run scan.
Coming soon