Indian–American Diaspora MCP Server

Agent-first, agent-only data layer for the Indian-American diaspora. Phase 1: Indian restaurants (USA). This repo is the walking skeleton from the architecture blueprint:

• MCP server exposing restaurant capabilities (get_indian_restaurants, get_restaurant_details, search_restaurants_by_text, find_unclaimed_restaurants, draft_claim_outreach, submit_correction), temple capabilities (get_indian_temples, get_temple_details, search_temples_by_text) and grocery capabilities (get_indian_groceries, get_grocery_details, search_groceries_by_text) professional capabilities (get_indian_professionals, …), salon capabilities (get_indian_salons, …), and a cross-vertical search_all that searches every vertical at once (each result tagged with its vertical).

• Data pipeline: scrape → raw → clean/enrich/score → approval queue → canonical table → versioning.

• One real scraper: OpenStreetMap Overpass (public, ODbL-licensed, no login, ToS-safe).

• Storage: PostgreSQL (+ pgvector for future embedding search) via Docker Compose.

Connect (live MCP server)

The hosted server is remote (no install) over streamable‑HTTP:

• Endpoint: https://namasteamerica.us/mcp (transport: streamable-http, auth: none, read‑only)

• Descriptor: https://namasteamerica.us/.well-known/mcp.json · Docs: https://namasteamerica.us/for-agents

{ "mcpServers": { "namaste-america": { "type": "streamable-http", "url": "https://namasteamerica.us/mcp" } } }

Tools (per vertical: restaurants, temples, groceries, professionals, salons, events, apparel, sweets, studios, services, community, legal, education, real estate, finance): get_indian_<category>, search_<category>_by_text, get_<category>_details, and a cross‑vertical search_all. Each result is JSON with address, geo, hours, contact, ratings, and languages.

Publishing to the official MCP registry: the manifest is server.json (namespace us.namasteamerica, verified via DNS). Install the CLI (mcp-publisher), authenticate the domain (mcp-publisher login dns --domain namasteamerica.us, then add the TXT record it prints), and run mcp-publisher publish.

Related MCP server: LandlordLaw MCP Server

Architecture (Phase 1)

 OSM Overpass ──▶ restaurant_raw ──▶ clean()+score() ──▶ approval_queue ──▶ restaurants
   (scraper)        (JSONB)            (pipeline)         (high-risk only)   (canonical)
                                                                                 │
                                                                                 ▼
                                                                        restaurant_versions
                                                                                 │
                                                                                 ▼
                                                                     MCP tools (FastMCP)

Low-risk new inserts are auto-applied (configurable); updates to claimed/featured listings are routed to the human approval queue.

Quick start

Requires Docker Desktop and Python 3.11+.

# 1. Start Postgres (with pgvector)
docker compose up -d

# 2. Create a virtualenv and install
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .

# 3. Configure env
copy .env.example .env   # default DATABASE_URL already points at the docker container

# 4. Initialise the schema
python -m indo_usa_mcp.cli init-db

# 5. Scrape one metro from OpenStreetMap, then process raw -> canonical
python -m indo_usa_mcp.cli scrape --metro bay_area
python -m indo_usa_mcp.cli process

# 6. Inspect
python -m indo_usa_mcp.cli stats

# 7. Run the MCP server (stdio transport)
python -m indo_usa_mcp.server

CLI

Monetization & delivery

• Featured listings (feature/unfeature) — a paid tier; effectively-featured rows (flagged and within their featured_until window) surface first in every tool result.

• Stripe checkout (optional) — claimed owners get a "Get Featured" button → Stripe Checkout → a signature-verified webhook (/stripe/webhook) auto-features them. Disabled until STRIPE_SECRET_KEY is set (then it's pay-per-sale, no monthly fee). See DEPLOY.md.

• Outreach email delivery — optional and zero-cost: set Gmail SMTP + an app password in .env (see .env.example) and the Outreach Agent will auto-send claim emails to restaurants that have a public email; otherwise it stays draft-only.

• Enrichment agent — strengthens the cultural "data moat" by inferring region/dietary tags from restaurant names (free, keyword-based).

Coverage spans 15 metros (Bay Area, NYC/NJ, Dallas, Houston, Chicago, LA, Seattle, Atlanta, Phoenix, Austin, Boston, Philadelphia, Raleigh, Detroit, Central NJ) plus an on-demand nationwide sweep (scrape --metro usa).

WhatsApp outreach is delivered as free click-to-send wa.me links (message pre-filled); true auto-send WhatsApp needs a paid API and is intentionally not used.

Trying it without a live scrape

python -m indo_usa_mcp.cli init-db
python -m indo_usa_mcp.cli seed     # 12 fictional restaurants across the 5 metros
python -m indo_usa_mcp.cli stats

Then exercise the MCP tools (e.g. get_indian_restaurants near the Bay Area, or search_restaurants_by_text "vegetarian dosa").

Search quality (for LLM agents)

Records are optimized for agent retrieval:

• Natural-language description per record (e.g. "Saffron House is a Gujarati Indian restaurant in Edison, NJ. Offers vegetarian, jain options. Price: $$.") — generated from the structured fields, returned to agents, and used as the text that gets embedded.

• Reverse-geocoded location — missing city/state are filled offline from coordinates (reverse_geocoder), so "near me / in " queries work for every record.

• Keyword + attribute tags (e.g. biryani, dosa, halal, delivery, takeout, outdoor-seating, wheelchair-accessible, wifi, cards-accepted, organic) — derived from names and rich OSM attributes; agents filter with tag=..., and they boost embedding recall + appear in the description ("Amenities: …").

• open_now — opening_hours are parsed into structured per-day intervals; each result carries an open_now flag (true/false/null) and the tools accept an open_now=true filter, enabling "what's open near me right now".

• Dedup-merge — the admin Quality view lists duplicate groups with a one-click merge (fills the keeper's empty fields, soft-deletes the rest).

search_*_by_text ranks by embedding cosine distance (pgvector <=>), falling back to trigram. Providers (EMBEDDING_PROVIDER):

• hashing (default) — feature-hashing, zero extra deps; lexical.

• fastembed (recommended, prod default) — real semantic embeddings via BAAI/bge-small-en-v1.5 (384-dim, ONNX, no torch). The Docker image includes it and the prod compose defaults EMBEDDING_PROVIDER=fastembed; after deploying run python -m indo_usa_mcp.cli enhance-data once to re-embed existing rows. Set EMBEDDING_PROVIDER=hashing if the VPS is RAM-constrained.

• sentence_transformers — all-MiniLM-L6-v2 (heavier, pulls torch).

• none — trigram only.

enhance-data (re)generates descriptions, fills geocoding, and re-embeds existing rows — run it after enabling a new embedder. New records get all of this automatically on ingest.

Outreach & claiming (blueprint §7)

# Draft claim messages + single-use claim links for unclaimed restaurants
python -m indo_usa_mcp.cli outreach --limit 10

# Later, an owner verifies their claim token (normally via the claim web page)
python -m indo_usa_mcp.cli verify-claim <token> --email owner@example.com

After claiming, owners get an edit page (/manage?token=...) to update their phone, hours, menu, price, dietary tags, etc. — changes go live immediately and are protected from scraper overwrites (scraper updates to claimed listings route to the approval queue). Stale listings not re-seen for 60 days are auto-deactivated (is_active=false) and reactivated if they reappear in a later scrape.

The Outreach Agent finds unclaimed restaurants (skipping anything with an open claim or contacted within OUTREACH_COOLDOWN_DAYS), creates a single-use claim token + link, and drafts an honest, opt-out-friendly message per restaurant. Messages are not auto-sent — delivery needs channel integrations, and chains / featured / high-value targets are flagged requires_human. Once an owner verifies, restaurants.is_claimed flips to true and future scraper updates to that listing are routed to the approval queue. Agents can drive this via the find_unclaimed_restaurants and draft_claim_outreach MCP tools.

Autonomous agents (blueprint §6)

Each agent wraps a pipeline step, is idempotent, and writes a full audit row to agent_runs (errors captured, never half-written canonical data).

python -m indo_usa_mcp.cli agents              # list agents
python -m indo_usa_mcp.cli agent scraper       # run one, audited
python -m indo_usa_mcp.cli agents-loop --once  # one scheduler pass (scrape→clean→monitor…)
python -m indo_usa_mcp.agents.scheduler        # long-lived scheduler (VPS worker)

Data sources

Two independent public scrapers feed the pipeline: osm_overpass (OpenStreetMap, ODbL) and wikidata (Wikidata SPARQL, CC0). Pick one with scrape --source <name>; the Scraper Agent runs both.

Phase 2: Temples vertical

An independent vertical (Hindu/Sikh/Jain places of worship) sharing the same pipeline, agents, embeddings and deployment — per the blueprint's "independent verticals, shared infra" principle. Own table (temples), own OSM scraper (amenity=place_of_worship + religion), own agents (temple_scraper, temple_cleaner), own MCP tools. Cultural enrichment infers deity (Venkateswara, Krishna, Lakshmi…) and region (Punjabi for Sikh, Gujarati for Swaminarayan, Telugu for Venkateswara…) from the name.

python -m indo_usa_mcp.cli temples-scrape --metro bay_area
python -m indo_usa_mcp.cli temples-process
python -m indo_usa_mcp.cli temples-stats
python -m indo_usa_mcp.cli temples-query --religion hindu --city Fremont
python -m indo_usa_mcp.cli temples-query --text "swaminarayan gujarati mandir"

Phase 2: Groceries vertical

Same recipe as temples — independent groceries table + OSM scraper (grocery-type shops whose name signals Indian groceries: Patel, India, Apna, Swad, Masala…) + own agents (grocery_scraper, grocery_cleaner) + 3 MCP tools. Infers store_type and region_tag.

python -m indo_usa_mcp.cli groceries-scrape --metro bay_area
python -m indo_usa_mcp.cli groceries-process
python -m indo_usa_mcp.cli groceries-query --city Fremont
python -m indo_usa_mcp.cli groceries-query --text "patel brothers indian grocery"

Admin dashboard, reporting & owner portal

The web app (:8080) now serves three audiences:

• Public/owner — /, /claim, /manage, /upgrade (existing)

• Admin (/admin/*, password-gated) — overview KPIs, cross-vertical data control (browse/search/edit/feature/deactivate/soft-delete), Geography drill-down (country → state → city, click to filter records), Quality control (records flagged for missing region/contact/geo/city + duplicate groups, with a one-click city/state normalizer), approvals & feedback queues, agents (last-run health, run-now, resolve alerts), Traffic (every MCP tool call logged: by tool / agent-client / day, plus most-shown listings), payments, and reports. Set ADMIN_PASSWORD to enable; blank disables it.

• Owners see a reach number per listing in the portal ("shown N times by AI assistants", from per-listing impression tracking) — a concrete hook for selling featured placement.

• Customer portal (/portal/*) — owners sign in via passwordless magic-link email and manage all their listings across verticals (edit, featured status, upgrade).

Daily report: the reporting agent computes a health + growth snapshot nightly into daily_reports, shows it on /admin/reports, and emails it (via SMTP) to REPORT_EMAIL. Run on demand with python -m indo_usa_mcp.cli report.

Security: enable HTTPS (the Caddy tls profile) and set a strong ADMIN_PASSWORD + random SECRET_KEY before exposing /admin publicly.

Phase 2: Professionals vertical (doctors/health)

Indian-American healthcare professionals — doctors, dentists, clinics, pharmacies — found from OSM healthcare amenities + an Indian-name signal (surnames / Ayurveda). Same recipe as the other verticals: own professionals table, scraper, agents (professional_scraper, professional_cleaner), 3 MCP tools, plus profession_type/speciality filters. Because name-matching is heuristic, each record carries a confidence_score and the admin Quality view helps curate false positives.

python -m indo_usa_mcp.cli professionals-scrape --metro bay_area
python -m indo_usa_mcp.cli professionals-process
python -m indo_usa_mcp.cli professionals-query --type dentist --city "San Jose"

Phase 2: Salons vertical (beauty / threading)

Indian beauty salons — eyebrow threading, henna/mehndi, hair, bridal — from OSM hairdresser/beauty shops with a strong South-Asian service/name signal. Same recipe; tags capture services (threading, henna, bridal…). Best OSM coverage of the verticals so far.

python -m indo_usa_mcp.cli salons-scrape --metro bay_area
python -m indo_usa_mcp.cli salons-process
python -m indo_usa_mcp.cli salons-query --tag threading --city Sunnyvale

Phase 2: Events vertical (automated, admin-approved)

Indian-American community events (festivals, garba, concerts, puja). Fully automated — an agent auto-discovers iCal feeds by scanning the websites of orgs already in the database (temples first) for calendar links, and agents ingest events from those feeds plus any in EVENT_ICAL_FEEDS. The cleaner auto-approves high-confidence events and queues the rest under Admin → Events for one-click approval (no manual entry). Past events are kept, not deleted — they're date-filtered out of upcoming results (get_indian_events returns upcoming by default; include_past=true for history) and an 18-month retention purges very old ones.

python -m indo_usa_mcp.cli events-discover     # scan org websites for iCal feeds
python -m indo_usa_mcp.cli events-scrape      # ingest configured + discovered iCal feeds
python -m indo_usa_mcp.cli events-process     # auto-approve / queue for admin
python -m indo_usa_mcp.cli events-approvals   # list pending
python -m indo_usa_mcp.cli events-query --city Edison --category garba

Connecting an MCP client

Point any MCP client (Claude Desktop, etc.) at:

{
  "mcpServers": {
    "indo-usa-diaspora": {
      "command": "python",
      "args": ["-m", "indo_usa_mcp.server"],
      "env": { "DATABASE_URL": "postgresql://diaspora:diaspora@localhost:5433/diaspora" }
    }
  }
}

Deployment (VPS, blueprint §10)

The full hosted stack — Postgres + MCP server (HTTP) + agent worker — is in docker-compose.prod.yml:

$env:POSTGRES_PASSWORD = "choose-a-strong-password"
docker compose -f docker-compose.prod.yml up -d --build

• server runs migrations (idempotent) then serves MCP over streamable-http on :8000.

• worker runs the agent scheduler (scrape → clean → monitor …); it waits for the schema before starting, so service start order doesn't matter.

• db stays on the internal network (not published).

Put a TLS-terminating reverse proxy (Caddy/nginx/Traefik) in front of :8000 for HTTPS, and add a firewall + scheduled pg_dump backups. The MCP endpoint is https://<host>/mcp.

Checking it with MCP Inspector

The standard visual way to exercise the tools:

# stdio (local)
npx @modelcontextprotocol/inspector .\.venv\Scripts\python.exe -m indo_usa_mcp.server

# or point the Inspector at a running HTTP server: http://localhost:8000/mcp

Guardrails honoured

• Public data only (OSM/ODbL), no login-required scraping, no personal data, rate-limited.

• Confidence scoring, soft deletes, full version history.

• Human approval for high-risk updates; is_featured is an explicit, visible flag.

Status / roadmap

Phase 1 is implemented and runs end-to-end: schema + versioning + approval queue, the pipeline, two scrapers (OSM + Wikidata), all 5 MCP tools, semantic (pgvector) search, outreach & claiming, six autonomous agents with a scheduler, seed fixtures, and a hosted deployment stack. Remaining "last mile": real outreach channel delivery, a claim web page, the Approval-Assistant + Feedback agents, monetization logic, and production hardening (HTTPS, backups). Each future vertical (temples, events, groceries…) gets its own table/scrapers/tools but shares this infra.