Overview
A remote MCP server that turns Claude into a personal assistant — tracks habits, tasks, finances, goals, documents, and memory with semantic search. Built on OAuth 2.0 + PKCE, deployed serverless on Vercel with 41 MCP tools across 6 domains.
Tech Stack
Next.js 16React 19TypeScriptSupabasePostgreSQLpgvectorMCP SDKZodOpenAI EmbeddingsTailwind CSSVercel
Key Features
41 MCP tools across Habits, Tasks, Finance, Goals, Documents, and Memory
OAuth 2.0 with PKCE (S256) for secure remote MCP client authentication
Hybrid semantic + full-text search on memories using pgvector and GIN indexes
Duplicate memory detection at 0.9 cosine similarity threshold
Document wallet with chunked embeddings (OpenAI text-embedding-ada-002, 1536 dims)
Habit streak tracking and completion analytics with IST-aware date handling
Spending categorization with uncategorized-transaction surfacing
Goal tracking with milestone sub-steps and progress reviews
Stateless serverless architecture — new MCP instance per request
SHA-256 hashed OAuth tokens, service-role Supabase access server-side
Dashboard UI for all 6 domains alongside MCP access
Challenges & Solutions
Stateless MCP server on Vercel serverless
MCP protocol typically assumes persistent connections, but Vercel functions can't hold state between requests. Built a fresh MCP server instance on every POST /api/mcp call — no module-level state, no singletons — so it scales horizontally with zero warm-up coupling.
OAuth 2.0 + PKCE for remote MCP clients
Claude.ai and other MCP clients require OAuth for remote servers — no API key auth. Implemented full /oauth/authorize, /token, /register, /revoke flow with PKCE S256, and SHA-256 hashed tokens in Supabase so raw tokens never hit the DB.
Recall quality on memory search
Pure vector search missed exact keyword hits like names, dates, and IDs. Built a pa_hybrid_search RPC combining pgvector cosine similarity with Postgres full-text search over a GIN-indexed search_vector column, giving the best of both worlds.
Sharing a Supabase project with memory-mcp without collisions
Sathi coexists with an earlier memory-mcp server on the same Supabase instance. Namespaced every Sathi table with a pa_ prefix (pa_memory_items, pa_memory_spaces, pa_memory_access_log) to keep schemas isolated and migrations conflict-free.
Timezone correctness for an India-based user
Streaks and daily filters kept going off-by-one because server default was UTC. Standardized everything on Asia/Kolkata via IST helpers in types/index.ts — all display, filtering, and habit-completion logic now uses IST consistently.
Architecture
Next.js 16 + Supabase (pgvector) + MCP SDK + OAuth 2.0 PKCE, deployed on Vercel serverless
Next Project