weight-mcp

A personal calorie & protein counter delivered as an MCP server for use inside claude.ai. You log meals from photos or text in a normal Claude chat; this server counts them, tracks weight, and renders an interactive dashboard (weight graph + recently eaten) right in the conversation.

Self-hosted, multi-user: one admin account lives in .env, and the admin registers further accounts from chat — each user sees only their own data. See SPEC.md for the rationale.

How it works

• MCP server (Streamable HTTP) exposes tools, a prompt, and an MCP Apps dashboard UI to claude.ai.

• In-process OAuth gate: claude.ai drives the full OAuth 2.1 + PKCE + Dynamic Client Registration flow; the only human step is a username + password form. The admin signs in as admin with the password from .env; everyone else signs in with an account the admin registered.

• SQLite stores accounts, weights, and food logs — one file under data/. Every data row is scoped to the account that logged it.

Related MCP server: fitbit-googlehealth-mcp

Tools

All of the above operate on the calling account's own data. The admin account additionally gets user management:

Configure

Copy .env.example to .env and set at least WEIGHT_MCP_PASSWORD (the admin account's password — the only account configured in .env) and WEIGHT_MCP_PUBLIC_BASE_URL. Non-admin accounts are not env config: the admin creates them from chat with register_user. Nutrition sources default to Open Food Facts, filtered to Germany; set WEIGHT_MCP_* to change region or enable USDA.

Goals are not env config — set them from chat with set_goals (they persist in the database). Two modes: floor (eat at least the target — the default, for under-eaters) and ceiling (stay under — for weight loss). Until you set your own, the default is 2600 kcal / 150 g protein, floor. Goals are per account, like all other data.

Accountability reporting is optional: set WEIGHT_MCP_TELEGRAM_BOT_TOKEN (a bot from @BotFather) and WEIGHT_MCP_TELEGRAM_CHAT_ID (your partner's chat — a human, group, or channel; numeric ID or @channelusername) to enable send_daily_report. The bot must be able to message that chat: the partner has to start the bot once (humans), or the bot must be a member/admin (groups/channels).

Run

Local (dev):

uv sync
cp .env.example .env   # then edit
uv run weight-mcp

Docker (local build):

docker compose up --build

For a real deployment you need public HTTPS (claude.ai connects from Anthropic's cloud, not your device). Copy docker-compose.template.yml, put the server behind a TLS reverse proxy, and set WEIGHT_MCP_PUBLIC_BASE_URL to that origin.

Add to claude.ai

Customize → Connectors → "Add custom connector" (on Team/Enterprise: Organization settings → Connectors → Add → Custom → Web), paste your server's base URL (https://<your-host> — the MCP endpoint and OAuth live at the origin root, so there is no path to append). claude.ai opens the OAuth page; sign in with your username and password (admin + the .env password for the admin, or credentials the admin registered for you). Done — start a chat and tell Claude what you ate.

Develop

uv run ruff check .
uv run mypy
uv run pytest