A calorie tracker that lives in your terminal — and in your pocket.
Claude Code (or claude.ai on your phone) is the app; a private git repo you own is the database.

> log: 2 rotis and a katori of dal

Logged to 2026-07-04 · lunch
  roti ×2 · 80 g (your roti, calibrated) · 240 kcal
  dal · 150 g (your katori) · 158 kcal
Day so far: 1,240 kcal · 52 g protein · on-plan

No server, no account, no subscription. You say what you ate (or paste a photo), the agent resolves it against real nutrition data, writes a JSON file into your repo, and commits it. In fifty years, when every app in this story is dead, your food history is still dated JSON files you can open in anything.

The problem

Mainstream calorie trackers fail in the same three ways:

1. The essentials are paywalled. Photo logging, macro breakdowns, and exporting your own data all sit behind a subscription.
2. Non-Western food data is garbage. The same roti ranges from 70 to 150 kcal across crowd-sourced duplicate entries. If you eat dal, sabzi, or anything your grandmother would recognize, the numbers are noise.
3. Your history is hostage. Years of logs live in someone else's cloud, readable only through their app, gone when they pivot or fold.

And underneath all three: logging through search-a-database-and-pick-from-a-dropdown is enough friction that most people quit within weeks.

What gohan does instead

• Natural language is the interface. "2 rotis and a katori of dal" needs no form and no barcode. An LLM is the best food parser ever built; gohan just points it at grounded data.
• Your data, your repo. Each user runs their own private copy. Nothing is hosted, shared, or metered. git log is your audit trail.
• Every number is traceable. Each logged item keeps your words verbatim, the resolved food, how the grams were derived, the nutrition source, and a confidence flag. When gohan estimates, it says so.
• Counting, not coaching. gohan records and reports. What you should eat is between you and your dietician. gohan gives you honest numbers to bring to that conversation.

Install

You need Claude Code and bun. The repo doubles as its own plugin marketplace:

/plugin marketplace add shiroyasha9/gohan
/plugin install gohan@gohan

Then make an empty directory for your food history, open Claude Code in it, and say:

set me up

The setup skill interviews you (timezone, day boundary, optional stats), scaffolds your private data repo, walks you through creating a private remote, and offers two optional imports:

• Indian food data: downloads and converts the INDB dataset (~1,000 Indian recipes) locally on your machine. Not eating Indian food? Skip it.
• USDA lookups: a free FDC API key covers everything else.

Updating later is /plugin update gohan. Your data repo is never touched.

Important

Your data repo is personal health data. Keep it private. Setup will never offer to create a public remote, and neither should you.

Use it from your phone (claude.ai)

The same tracker works in the Claude app with no server and no extra install — the plugin provides the skills, and GitHub's own MCP server gives Claude read/write access to your data repo:

1. Install the plugin: claude.ai → Settings → Plugins → Add → Add marketplace → Add from a repository → shiroyasha9/gohan → Install.
2. Create a GitHub OAuth App (one-time, 2 minutes; GitHub doesn't support automatic client registration, so each user brings their own): GitHub → Settings → Developer settings → OAuth Apps → New. Homepage https://claude.ai, authorization callback https://claude.ai/api/mcp/auth_callback. Register, then generate a client secret.
3. Add the connector: claude.ai → Settings → Connectors → Add custom connector → URL https://api.githubcopilot.com/mcp/x/repos, paste your OAuth App's client ID and secret → Connect → authorize on GitHub.
4. Name your data repo gohan-data so skills find it on their own — or just tell the chat which repo it is, once per conversation. No repo yet? Say "set me up" and setup scaffolds one.
5. Verify in a fresh chat on the mobile app: log: 1 banana → a log: commit should land on your repo's main branch.

Caveats (as of 2026-07-11):

• Use the mobile app for now. claude.ai web has an open bug (anthropics/claude-ai-mcp#476) where custom connector tools never reach the model; the phone app works.
• Scope: a GitHub OAuth App grant covers every repo your account can push to, not just gohan-data. If that bothers you, a dedicated GitHub account owning only the data repo narrows it completely.
• Desktop stays the fully-tooled surface: on the phone, new foods resolve via web search with cited sources instead of the local INDB/USDA pipeline, and long summaries or bulk recomputes will point you back to a desktop session.

Daily use

You talk; skills route:

You say	What happens
log: 2 rotis and dal / a meal photo	log-meal writes today's day-file, commits, shows totals
weight: 81.4 / gym: legs, 60min	log-body records weight, workouts, treatment notes
summary / how was this week	summarize shows totals, adherence %, weight trend
my katori weighs 150g	calibrate maps that unit to real grams, permanently
share a photo of your diet plan	import-diet-chart versions it as data
(a new food shows up)	resolve-food finds grounded numbers, caches them
(a cached food gets corrected)	recompute silently fixes every affected past day

Every write ends with a commit and push, so phone and laptop sessions never diverge.

How your data looks

data/
├── days/2026/07/2026-07-04.json   one file per day: meals, weight, workouts
├── foods.json                     personal food cache, per-100g, source-flagged
├── measures.json                  your utensils, weighed once ("my katori = 150 g")
├── diet-chart/                    optional: your nutritionist's plan, versioned
└── profile.json                   timezone, day boundary, stats

Each food is resolved once (nutrition database first, then the USDA API, then a clearly flagged estimate) and becomes canon for every repeat log. Trends stay meaningful even when an absolute value is somewhat off, because the same dal is always the same dal.

A logged item is a full resolution trail:

{
  "input": "1 katori dal",
  "food": "dal (home)",
  "qty": { "amount": 1, "unit": "katori", "grams": 150, "basis": "measures" },
  "nutrients": { "kcal": 158, "protein": 9, "carbs": 21, "fat": 4, "fiber": 6 },
  "source": "indb:DAL_TADKA",
  "confidence": "high"
}

Conventions worth knowing

• Wake-day: entries before your configured cutoff (default 04:00) count toward the previous day. A midnight snack belongs to the day you were awake for.
• Confidence per item: high means a cached food with weighed or calibrated grams, medium means estimated grams, low means an estimated food.
• Burn is display-only: workout calories are tracked for consistency, but they are never netted against intake. Fictional burn numbers don't get to justify extra eating.
• Adherence is derived, not invented: with a diet chart imported, every meal gets an on-plan / variation / cheat label, and summaries compare intake against your nutritionist's targets rather than numbers gohan made up.

Development

This repo is the dev workspace; plugin/ is what ships (skills, prebuilt bun bundles, schemas copy, session hook). After changing apps/scripts or packages/core, run bun run build:plugin and commit the regenerated bundles. CI fails on drift.

Dev loop: claude --plugin-dir ./plugin from a scratch data directory, /reload-plugins after edits. Quality gates: bun test, bun run typecheck, bun run check. packages/core/src/schemas.ts is the single source of truth for every data shape; design docs live in specs/.

License

MIT