mealforge
Enables AI-driven weekly meal planning with structured recipes, cooking steps, and grocery lists.
README
mealforge
AI-planned weekly meals, in a real app. Plan your week conversationally with any MCP-capable AI chat (LibreChat, Claude, and friends) — the model pushes the finished plan here, where it becomes recipe cards, a step-by-step cook mode, and a grocery list you check off in the store.
Meal-planning apps limit you to their recipe catalog. mealforge has no catalog: every recipe is generated in conversation, tailored to your household, constraints, and whatever sounds good this week. The app is the structured, shoppable, cookable home for what you and the model decide together.
How it works
┌────────────┐ MCP (streamable http) ┌────────────────────────────┐
│ your AI │ ───────────────────────▶│ mealforge │
│ chat │ push_meal_plan │ • weekly plan view │
│ (LibreChat,│ search_recipes │ • recipes + cook mode │
│ etc.) │ list_favorites … │ • grocery list (derived) │
└────────────┘ │ • favorites + history │
└────────────────────────────┘
- Plan in chat. "Lots of crock pot this week, salmon once, use the pork shoulder in the freezer." Iterate until it's right.
- The model pushes the final plan (
push_meal_plan): structured recipes — ingredients with quantities, units, and store sections — plus markdown cooking steps. - mealforge derives the grocery list automatically: ingredients aggregated across recipes ("2 cups" + "1 cup" → "3 cups"), grouped by store section, checkable while you shop.
- Favorite what you loved. The model can recall favorites and past recipes next week ("give me one of my favorites") and avoids repeating recent meals.
Self-hosting
mkdir mealforge && cd mealforge
curl -O https://raw.githubusercontent.com/loosewiredev/mealforge/main/docker-compose.yml
docker compose up -d
The web app and MCP endpoint are now on port 8090:
- Web UI:
http://<host>:8090 - MCP endpoint:
http://<host>:8090/mcp
Data lives in a single SQLite file under ./data/ — back that folder up and you've backed up everything.
Security note: mealforge has no built-in authentication. It is designed for a household on a private network — put it on your Tailscale/VPN, or behind a reverse proxy that does auth. Do not expose it to the public internet as-is.
Set APP_URL (in the compose file or a .env next to it) to the URL your household uses, e.g. APP_URL=https://mealforge.your-tailnet.ts.net — the MCP tools return it so the model can link you to the app.
Connecting an MCP client
LibreChat
librechat.yaml:
mcpServers:
mealforge:
type: streamable-http
url: http://<host>:8090/mcp
timeout: 60000
If mealforge runs on a private IP, also exempt it from LibreChat's SSRF guard:
mcpSettings:
allowedAddresses:
- "<host>:8090"
Then restart the LibreChat container and enable the mealforge tools for your agent/endpoint.
Other clients
Any client that speaks MCP over streamable HTTP works — point it at http://<host>:8090/mcp. No auth headers are required (see the security note above).
Give your agent the meal-planning skill
The repo ships a ready-made agent skill at skills/weekly-meal-planning/SKILL.md. It teaches an agent the whole workflow: gather history first, draft the week in conversation, publish only on explicit command via the reliable two-step flow (create_recipe per recipe, then one push_meal_plan with recipeIds), keep ingredient data grocery-list-clean, and recover from validation errors. Strongly recommended — especially with smaller models.
- Claude Code / Agent Skills: copy the
skills/weekly-meal-planning/directory into your skills folder (e.g.~/.claude/skills/). - LibreChat: paste the body of
SKILL.mdinto your meal-planning agent's instructions (or attach it as an agent skill/file). - Anything else: it's plain markdown — hand it to your agent however that client takes instructions.
MCP tools
| Tool | Purpose |
|---|---|
create_recipe |
Save one recipe (flat payload) and get back a recipeId — the reliable first step before push_meal_plan. |
push_meal_plan |
Push a finalized week (new recipes and/or recipeId reuses). Re-pushing the same weekStart revises the week; checked-off grocery items that didn't change stay checked. |
get_recent_meal_plans |
Recent weeks with meal titles — for repeat-avoidance. |
get_meal_plan_for_week |
The plan for a specific week, if any. |
list_favorites |
Recipes the household has favorited in the UI. |
search_recipes |
Search past recipes by title, tag, or ingredient. |
get_recipe |
Full recipe (ingredients + steps) by id. |
Development
pnpm + Nx monorepo: apps/api (Hono + tRPC + Drizzle + SQLite, serves the MCP endpoint and the built web app), apps/web (React + Vite + TanStack Router + Tailwind).
pnpm install
pnpm nx run-many -t test,build # tests + builds
pnpm --filter @mealforge/api dev # API on :3000
pnpm --filter @mealforge/web dev # web on :5173 (proxies /trpc to :3000)
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。