AI Usage Dashboard

Turn your ChatGPT export into a navigable personal-story dashboard. Runs locally on your own machine, your data never leaves your laptop.

What it does

The dashboard reads your ChatGPT data export and produces four views over your last few years of conversation:

• Overview — totals, streaks, gaps
• Rhythms — when, what model, how much, how long
• Timeline — monthly narratives + era detection
• Topics — embeddings, UMAP scatter, named clusters

It's not a chat browser. It's a tool for looking at the shape of three years of asking AI questions — what you were working on, learning, wrestling with — and seeing patterns you wouldn't notice scrolling through them one at a time.

Highlights

A monthly narrative anchored to the conversations underneath

Claude Sonnet reads every conversation summary from a given month and writes a paragraph capturing what you were actually working on. Highlight cards link straight to the originals.

Watch the model migration

Stacked-area chart of every assistant message by model family — the gpt-4 era, the gpt-4o ramp, the move to gpt-5 thinking.

When you actually use it

Hour × day-of-week heatmap surfaces your real rhythm — the late-night research bursts, the consistent morning queries, the weekend gaps.

Conversation length, voice usage

Topic clusters

Every conversation embedded with OpenAI text-embedding-3-small, projected to 2D with UMAP, k-means clustered, and labeled by Claude Sonnet. The axes mean nothing — only distance matters.

Click any cluster to focus the scatter and explore its representative conversations.

Quickstart

1. Request your export from ChatGPT: Settings → Data Controls → Export → wait for the email.
2. Clone this repo:

   git clone https://github.com/YOUR_USERNAME/ai-usage-dashboard
   cd ai-usage-dashboard

3. Drop the Conversations__*.zip into OpenAI-export/User Online Activity/.
4. Set up env:

   cp .env.local.example .env.local
   # Edit .env.local — add your Anthropic + OpenAI keys, optional DASHBOARD_PASSWORD

5. Install + run the pipeline:

   npm install
   npm run ingest                       # ~30s, free
   ANALYZE_LIMIT=20 npm run analyze     # validate prompt quality first
   npm run analyze                      # full run, ~20 min, ~$1–2
   npm run normalize                    # canonical topic buckets
   npm run embed                        # ~3 min, ~$0.05
   npm run narratives                   # ~10 min, ~$2–5

6. Start the app:

   npm run dev

7. Open http://localhost:3000.

Cost

Step	Cost	Provider
ingest	$0	local
analyze (1,200 convos)	~$1–2	Anthropic Claude Haiku 4.5
embed (1,200 convos)	~$0.05	OpenAI text-embedding-3-small
narratives (~36 months)	~$2–5	Anthropic Claude Sonnet 4.6
Total	~$5–10 one-time

All LLM responses are cached on disk by content hash, so re-running is free for already-processed conversations.

How it works

A four-stage offline pipeline writes to a local SQLite file, and a Next.js app reads from it. Nothing reaches into the LLM on a page load.

ChatGPT export → ingest → analyze → embed → narratives → SQLite → Next.js
                  30s      20m       3m       10m

See docs/architecture.md for a deeper tour of the pipeline, schema, and design choices.

Tech

Next.js 15 (App Router) · TypeScript · Tailwind · SQLite (better-sqlite3) · Observable Plot · Anthropic SDK · OpenAI SDK · umap-js

Privacy

Your conversations contain a lot. By default the dashboard:

• Runs only on your machine. No network calls except to Anthropic/OpenAI during the pipeline.
• Gates access with a single-password middleware when DASHBOARD_PASSWORD is set in .env.local (httpOnly cookie, no session storage).
• Never commits data. OpenAI-export/ and data/ are both .gitignored.

If you deploy somewhere (Vercel, Fly), set DASHBOARD_PASSWORD to keep the dashboard private.

License

MIT — see LICENSE.