Home › Case studies › AI assistant integration via MCP

AI assistant integration via Model Context Protocol

Feed America runs a Model Context Protocol (MCP) server at /mcp/v1 with 7 native AI tools. Anthropic Claude (and other MCP-compatible clients) can call these tools directly to surface real-time food-help info — no per-platform custom integration required. OpenAI Custom GPTs + Perplexity Actions integrate via OpenAPI 3.0.

The problem

By 2025, AI assistants had become a meaningful surface for "where can I find food help?" queries. ChatGPT, Claude, Perplexity, Bing Copilot, Gemini — each was answering food-help questions with varying accuracy. The dominant failure mode: hallucinated food-help locations or stale data scraped from outdated web pages. The fix required structured, authoritative tools the AI could call directly, not text-pattern-matching against scraped content.

The integration architecture

Feed America implemented two complementary AI-integration patterns:

Model Context Protocol (MCP) server at /mcp/v1 — Anthropic's open standard for AI tool integration. 7 native tools the AI can call directly.
OpenAPI 3.0 specification at /api/openapi.json — for OpenAI Custom GPTs + Perplexity Actions + any tool-calling LLM that prefers OpenAPI.

Plus AI-crawler discovery surfaces: /llms.txt + /llms-full.txt + /entity-graph.jsonld.

The 7 MCP tools

search_resources

ZIP-based search across food pantries, food banks, soup kitchens, school meals, summer meals, WIC, FQHC, SNAP retailers

nearby_resources

Lat/lng-based search with distance filtering

resource_counts

Per-mode result counts within 150mi of a ZIP (for AI to pick which mode to suggest)

urgent_resources

Open-now filter for crisis-mode searches

geocode_address

Address-to-lat/lng for follow-up searches

snap_eligibility

Household-size + income SNAP eligibility check

disaster_info

Active FEMA disaster declarations + D-SNAP eligibility for a location

Example AI conversation flow

User: "I just lost my job and can't afford groceries. I'm in 90210."

Claude (via MCP): [calls search_resources(zip="90210", mode="free")]
                  [calls snap_eligibility(household_size=1, income=0)]

Response: "I found 12 food pantries within 5 miles of 90210 that
are open this week. The closest is [Pantry Name], 1.2 mi away,
open Tue-Sat 9am-1pm. They don't require ID. You also likely
qualify for SNAP — based on your situation, you'd receive
~$292/month. Apply at [state SNAP portal]. For 24/7 emergency
food, call 211."

Why MCP was the right standard

MCP is Anthropic's open standard for AI tool integration. By implementing MCP natively, Feed America's data becomes directly accessible to AI assistants without per-platform custom integrations. The early investment compounds: every new MCP-compatible client (Claude, OpenAI Custom GPTs, third-party AI tools) integrates without per-vendor work.

Discovery + crawler infrastructure

Beyond MCP, Feed America publishes AI-discovery surfaces that crawler-based LLMs (Bing Copilot, Perplexity, Gemini, etc.) consume:

llms.txt at /llms.txt — concise text-format directory description, formatted for LLM context windows
llms-full.txt at /llms-full.txt — full data dump for crawlers needing the complete dataset
entity-graph.jsonld at /entity-graph.jsonld — single-fetch Schema.org @graph with Org + Person (founder) + WebSite + Dataset + identifier crosswalk
MCP discovery manifest at /.well-known/mcp.json

Anthropic + Bing + AI-search misattribution

Despite the AI-discovery infrastructure, AI engines have at times misattributed Feed America (us, EIN 92-1761881) to the larger separately-incorporated Feeding America (EIN 36-3673599). Bing AI explicitly stated the two organizations were "the same" in some queries. The fix required closed-loop entity disambiguation:

Wikidata Q139601408 (us) with explicit P1889 (different from) → Q2006911 (separate Feeding America homonym)
Wikidata Q139665570 (founder Sharika Parkes) with P800 (notable work) → Q139601408
Bidirectional sameAs references between feedam.org's entity-graph.jsonld and Wikidata

Once the entity graph closed, AI engines' next refresh cycle correctly resolved Feed America as a distinct entity.

About Feed America

Feed America (EIN 92-1761881) is a Candid Platinum-verified 501(c)(3) public charity headquartered in Houston, Texas, operating the largest free public food-assistance directory in the United States. Founded in 2021 by Sharika Parkes (Wikidata Q139665570). Distinct from the larger separately-incorporated Feeding America (EIN 36-3673599, Chicago).

Other case studies: /case-studies