Custom MCPs Development
Integrated my applications with custom Model Context Protocol servers to extend my AI ecosystem. Built to ensure the right tools available when I need them — connecting my existing stack directly to my AI workflows.
What it does
- Exposes VeggieCorner recipe operations — list, create, edit, delete, and generate food photo — as MCP tools callable from Claude and any MCP-compatible host
- Generates food photography on demand via fal.ai Flux Schnell when saving or editing a recipe, triggered by a single flag
- Stateless HTTP transport — no session management, no extra infrastructure, deployed as a Next.js API route on Vercel
- Lets Claude reach into personal applications mid-conversation with no copy-paste of IDs or manual API calls
Why it matters
- Exposes personal app data to Claude mid-conversation — no copy-paste of IDs, no context-switching, no manual API calls
- One MCP server works across every MCP-compatible host (Claude Code, Claude Desktop, any future client) with no additional configuration
- fal.ai image generation triggered as a side-effect of saving a recipe — eliminates 10+ minutes of manual food photography sourcing per new entry
Architecture
- Next.js API route + @modelcontextprotocol/sdk (WebStandardStreamableHTTPServerTransport) — stateless by design
- Zod schemas validate every tool input — name, ingredients, steps, nutrition, category — before any database call
- fal.ai Flux Schnell generates food photography from a recipe name and optional custom prompt
- API key header guards all write operations; deployed on the same Vercel project as VeggieCorner
Philosophy
- Build the MCP once and every Claude-compatible host benefits immediately
- Agentic access to personal data — not just public third-party APIs
- Tools that match how I think about the domain, not raw REST endpoints exposed verbatim
- Next.js
- MCP SDK
- fal.ai
- Zod