claudegeist

The zeitgeist of your Claude Code sessions. A time-travelling word cloud of everything you've been thinking out loud to a language model at 2am.

You talk to Claude more than you talk to most people. Every prompt is a tiny confession of what you were stuck on, obsessed with, or quietly building. claudegeist reads all of it back, buckets it by time, and plays it forward so you can watch your own preoccupations rise, crest, and fade.

It's a word cloud. It's also a mirror. Mostly it's a word cloud.

        last week              this week             next sprint, probably
      ┌───────────┐          ┌───────────┐          ┌───────────┐
      │  migrate  │   ──▶    │  segfault │   ──▶    │  rewrite  │
      │  deadline │          │   WHY     │          │  in rust  │
      └───────────┘          └───────────┘          └───────────┘

---

Install

macOS, one line, no cargo, no brew, no excuses:

curl -fsSL https://raw.githubusercontent.com/sebasv/claudegeist/main/install.sh | bash

Grabs a universal (Intel + Apple Silicon) binary from the latest release and drops it in /usr/local/bin. Then:

claudegeist

It scans ~/.claude/projects, builds your timeline, and opens the cloud at http://127.0.0.1:8080. That's the whole onboarding.

Building from source? cargo run --release -- wherever you'd type claudegeist.

---

How to read your own brain

Three knobs, top of the screen. Twist them while it plays.

knob	options	what you're really looking at
source	prompts · assistant	your words / Claude's replies
metric	distinctive · frequency	what made each moment different / what you said most
bucket	daily · weekly · sprint	how zoomed-in time is

One of these deserves a footnote:

• distinctive (tf-idf) is the good one. Raw frequency just shows the, file, function forever — the vocabulary of work, identical every week. Distinctive surfaces what was new: the library you discovered Tuesday, the bug that ate Thursday. Play it on distinctive and you get a story. Play it on frequency and you get a screensaver.

Hit play. Words don't snap between weeks — they swell and shrink continuously, in place, so it actually feels like time passing instead of a slideshow. Drag the scrubber to jump. Pick your speed.

---

How it works

~/.claude/projects/**/*.jsonl  →  claudegeist  →  http://127.0.0.1:8080
        (your logs)              (one binary)       (the whole app)

1. Read. Every session is a JSONL log on your disk. claudegeist walks them all, pulls the text out of each message, and sorts it into two channels (your prompts and Claude's replies).
2. Count. Tokenise, drop stopwords and code-noise, tally per day.
3. Roll up. The daily counts collapse into weekly and sprint buckets, and each bucket gets scored two ways (raw count + tf-idf distinctiveness).
4. Serve. The viewer (web/index.html) is baked into the binary and the data lives in memory — so the compiled claudegeist is the entire product. Nothing to deploy, nothing to host, no node_modules in sight.

Rust + a stdlib HTTP server for the backend; d3 + d3-cloud for the cloud, with a stable layout (words keep their spot) and per-frame size interpolation (so they glide rather than jump).

Usage cheatsheet

claudegeist                       # scan ~/.claude/projects, serve at :8080
claudegeist /path/to/logs         # point it at a different logs dir
claudegeist -b daily              # open on daily buckets (default: weekly; also: sprint)
claudegeist /path/to/logs out.js  # write the data to a file instead of serving

---

Your secrets stay yours

It runs entirely on your machine. It binds to 127.0.0.1, holds everything in memory, and writes nothing by default — your half-formed 3am questions never leave the binary. The repo is public, but it ships code, not data: the generated buckets.js is gitignored on purpose. Your therapist's job is safe.

---

A small existential note

You can't see your own attention. It's spent moment to moment and then it's gone, with no record of where it went. This is one of the few records you have: not a to-do list of what you meant to care about, but a trace of what actually occupied you, week by week, in your own words.

Run it once a quarter. It's a surprisingly honest performance review.

---

_{Releases are cut by pushing a v* tag — CI builds the universal binary and publishes it, which is what the installer fetches. Tune TOP_N and the stopword list in src/main.rs.}