USDA FoodData Central MCP Server

Model Context Protocol (MCP) server that exposes USDA FoodData Central search and lookup tools. Plug it into Codex CLI, Claude Desktop, or any MCP-aware client to explore nutrition data without writing HTTP calls by hand.

Last README sync: base commit 8f49a48 (update after next commit).

Highlights

• Four ready-to-use tools wrapping FoodData Central search, single-record lookup, bulk lookup, and paginated listing.
• Strict validation with Zod schemas for inputs and outputs so LLMs can rely on structured results.
• Cursor-aware previews let you dry-run calls, request compact summaries, and opt into raw payloads only when needed.
• Lean nutrient lookups surface per-100 g calories, macros, saturated fat, and fiber through focused tools that accept just an FDC ID.
• Resilient HTTP client with throttling, timeouts, and exponential backoff retries for USDA rate limits.
• Built-in environment resource that describes the server configuration from inside your MCP client.

Requirements

• Node.js 18.19 or newer (Claude Desktop’s bundled Node works).
• USDA FoodData Central API key — request one at https://fdc.nal.usda.gov/api-key-signup.html and provide it via USDA_API_KEY.

Quick Start

git clone <repo-url>
cd USDA-mcp
npm install

cp .env.example .env
echo "USDA_API_KEY=your-key" >> .env

npm run start   # runs via tsx with stdio transport

The server exits immediately if USDA_API_KEY is missing or blank. When you hand the server off to an MCP client, have that client supply the variable instead of relying on .env.

To run the compiled CLI (needed for Codex autostart):

npm run build
npx usda-mcp            # assumes USDA_API_KEY is exported or supplied by the client
# Optional: install the CLI globally so `usda-mcp` is on your PATH
# npm install --global .
# (or run `npm link` inside the repo)

Configuration

Environment variables read at startup:

| Variable | Required | Default | Purpose | | ------------------- | -------- | ------- | --------------------------------------------------------------------------- | | USDA_API_KEY | Yes | — | FoodData Central API key; server exits if unset. | | USDA_API_BASE_URL | No | https://api.nal.usda.gov/fdc/v1/ | Override when routing through a proxy or staging host. |

You can provide these through .env, your shell, or the MCP client configuration. Use the config://usda-fooddata/environment resource to inspect the active settings from inside the client.

Running Under MCP Clients

Most MCP clients let you attach environment variables directly to a server definition. Provide USDA_API_KEY there so the USDA server runs regardless of your working directory. Only set USDA_API_BASE_URL when you need to hit a non-default endpoint.

Codex CLI (~/.config/codex/config.toml)

Codex profiles accept per-server environment variables in TOML (Codex configuration guide).

Use whichever command style matches your setup:

• command = "usda-mcp" if you ran npm install --global . (or npm link) so the CLI is on your PATH.
• command = "node" with args = ["/absolute/path/to/dist/server.js"] if you prefer not to install the package globally.

experimental_use_rmcp_client = true

[mcp_servers.usda_fooddata]
command = "usda-mcp"
startup_timeout_sec = 20
tool_timeout_sec = 60
# If the CLI is not on your PATH:
# command = "node"
# args = ["/Users/you/projects/USDA-mcp/dist/server.js"]

[mcp_servers.usda_fooddata.env]
USDA_API_KEY = "your-fooddata-central-key"
# Optional override if you proxy the API:
# USDA_API_BASE_URL = "https://api.nal.usda.gov/fdc/v1/"

Claude Desktop (claude_desktop_config.json)

Claude Desktop reads server definitions from ~/Library/Application Support/Claude/claude_desktop_config.json on macOS (see the MCP introductory docs for the format). Point to your built entry point and inject the key via env.

{
  "mcpServers": {
    "usda": {
      "command": "node",
      "args": ["/Users/you/projects/USDA-mcp/dist/server.js"],
      "env": {
        "USDA_API_KEY": "your-fooddata-central-key"
        // "USDA_API_BASE_URL": "https://api.nal.usda.gov/fdc/v1/"
      }
    }
  }
}

Re-run npm run build whenever you change the server so dist/server.js stays in sync.

Cursor IDE (~/.cursor/mcp.json)

Cursor keeps MCP definitions in ~/.cursor/mcp.json. Any server listed under mcpServers can set env (many server READMEs, including Yandex Search, use the same layout).

{
  "mcpServers": {
    "usda-fooddata": {
      "type": "stdio",
      "command": "node",
      "args": ["/Users/you/projects/USDA-mcp/dist/server.js"],
      "env": {
        "USDA_API_KEY": "your-fooddata-central-key"
      },
      "startupTimeoutMs": 20000,
      "toolTimeoutMs": 60000
    }
  }
}

Claude Code (settings.json)

{
  "mcpServers": {
    "usda-fooddata": {
      "command": "node",
      "args": ["/Users/you/projects/USDA-mcp/dist/server.js"],
      "env": {
        "USDA_API_KEY": "your-fooddata-central-key"
      },
      "timeout": 20000
    }
  }
}

Gemini CLI (settings.json)

Gemini CLI merges MCP servers from system, user, and workspace settings (Gemini CLI configuration).

{
  "mcpServers": {
    "usda-fooddata": {
      "command": "node",
      "args": ["/Users/you/projects/USDA-mcp/dist/server.js"],
      "env": {
        "USDA_API_KEY": "your-fooddata-central-key"
      },
      "timeout": 30000
    }
  },
  "mcp": {
    "allowed": ["usda-fooddata"]
  }
}

Tools

All tools return a plain-text summary plus a structuredContent payload with a summary object, compact previews, and (when requested) the raw USDA response. Use the preview and dry-run switches to conserve context until you know you need the full payload.

• search-foods – Full-text search that only surfaces the food description, optional brand/data type, and fdcId so agents can pick an entry without excessive detail. Filters, cursor pagination, sort controls, and dry-run previews help shrink context impact.
• get-food – Fetch a single FoodData Central (FDC) record by ID with optional format and nutrients filters. Requests default to the faster USDA “abridged” view; the summary highlights macros (when present) and any notable gaps in the response. When USDA retires a known legacy identifier (e.g., SR Legacy 4053 for olive oil), the tool automatically substitutes the documented replacement (Foundation 748608 in this case) and adds a note so you know why the ID changed.
• get_macros – Return per-100 g calories, protein, fat, and carbohydrates for a single FDC entry with structured nutrient metadata. The helper first issues an abridged request scoped to the macro nutrient IDs, escalates to format=full, retries again without any nutrient filter, and finally consults labelNutrients (including Foundation-style labels such as Energy, Total fat (NLEA), and Total carbohydrate (NLEA)). If a Foundation record still withholds even one macro after that sequence, the tool now errors with guidance so you can pick a different FDC entry or compute the numbers manually (e.g., 1 g fat ≈ 9 kcal).
• get_micros – Surfaces per-100 g vitamins and minerals (Calcium, Iron, Potassium, Sodium, Magnesium, Zinc, Vitamins A/C/D/E/K, Folate, Vitamins B6/B12). The helper runs through the same nutrient escalation/label parsing path, so most USDA foods return micronutrients without extra work.
• get_macro_micros – Combines both panels so you can fetch calories, macros, and the vitamin/mineral set in one call (useful when you would otherwise chain get_macros + get_micros back-to-back).
• get_fats, get_protein, get_carbs, get_kcal, get_satfats, get_fiber – Single-nutrient lookups that emit just the requested per-100 g value (or note that it is unavailable) to keep tool output distinct.
• get-foods – Bulk lookup for up to 50 FDC IDs in one call. Supports previewOnly, includeRaw, sampleSize, and estimateOnly so you can review lightweight previews before retrieving the full objects, defaults to the faster USDA “abridged” format, and flags any requested IDs the USDA API omits. Known legacy → replacement mappings (e.g., 4053 → 748608) are applied automatically and noted in the summary.
• list-foods – Deterministic paginated listing that accepts optional filters (data types, brand owner), cursor-based pagination, sort, and the same preview/dry-run switches as search-foods. The summary returns the next cursor only when another page is likely available.

Handling Foundation entries without USDA macros

Foundation datasets sometimes omit energy, protein, or carbohydrate rows entirely (pure oils often only list lipid chemistry). The macro helper therefore:

1. Requests the abridged payload scoped to the calorie/protein/fat/carbohydrate nutrient IDs.
2. Falls back to format=full so USDA can include “analysis only” nutrients.
3. Retries the abridged call without nutrient filters so the API can decide which nutrients to include.
4. Reads labelNutrients, including Foundation label names such as Energy (kcal), Total fat (NLEA), Protein (NLEA), and Total carbohydrate (NLEA) even when USDA exposes those values only under display-friendly keys.

If any macro is still missing and the entry’s dataType is Foundation, get_macros stops with an error that lists the missing fields and suggests either switching to a record that publishes macros (e.g., SR Legacy or Survey entries) or deriving them yourself. A quick rule of thumb: calories ≈ (fat_g * 9) + (protein_g * 4) + (carbs_g * 4). Some oils (including FDC 748608) still omit USDA-provided calories/protein/carbs entirely across abridged/full/label payloads—this is a USDA database gap, not an MCP parsing bug—so keep the manual derivation handy for edge cases that never expose those fields.

Need both macro and micronutrient panels simultaneously? Call get_macro_micros to collapse the workflow into a single USDA request. It enforces the same Foundation guard for missing macros, surfaces micronutrient gaps in the summary, and automatically falls back to an unfiltered USDA fetch if the combined nutrient filter list exceeds USDA’s request limit (so you still get data, just with a slightly larger payload).

Micronutrient coverage (get_micros)

USDA keeps dozens of micronutrients in FoodData Central. The get_micros tool focuses on the vitamins and minerals that appear on standard nutrition labels so agents can request them in one call. Each value is per 100 g and travels through the same abridged/full/label fallback path used by get_macros.

| Nutrient | Unit | USDA nutrient IDs (abridged/full) | Label aliases | | --- | --- | --- | --- | | Calcium | mg | 1087, 301 | calcium | | Iron | mg | 1089, 303 | iron | | Potassium | mg | 1092, 306 | potassium | | Sodium | mg | 1093, 307 | sodium | | Magnesium | mg | 1090, 304 | magnesium | | Zinc | mg | 1095, 309 | zinc | | Vitamin A (RAE) | mcg | 1104, 318 | vitaminA, vitamin a | | Vitamin C | mg | 1162, 401 | vitaminC, vitamin c | | Vitamin D (D2 + D3) | mcg | 1114, 324, 328 | vitaminD, vitamin d | | Vitamin E (alpha-tocopherol) | mg | 1109, 323 | vitaminE, vitamin e | | Vitamin K (phylloquinone) | mcg | 1185, 430 | vitaminK, vitamin k | | Folate, total | mcg | 1186, 417 | folate | | Vitamin B6 | mg | 1175, 415 | vitaminB6, vitamin b-6 | | Vitamin B12 | mcg | 1178, 418 | vitaminB12, vitamin b-12 |

If a micronutrient is absent from both the nutrient list and labelNutrients, the response lists it under summary.notes so downstream automation can decide whether to fall back to manual data.

get_macro_micros returns this entire table plus the macro set in one result, so you can minimize USDA round-trips when you know you need the full panel.

Resources

• config://usda-fooddata/environment – Markdown overview showing the active base URL, whether a key is detected, retry/throttle policies, and guidance for overrides.
• usda_rebuild_progress.md – Workspace log for ingredient coverage. Record every “no USDA equivalent” decision (for example, ground sumac lacks a generic record, so we either keep the branded FDC 2630657 macros or stay with the manual rollup) to prevent repeated lookups.

Operational Notes

• Requests time out after 30 seconds, use up to two retries with jittered exponential backoff on HTTP 429 or 5xx, and throttle to one concurrent call with ≥400 ms spacing so we stay under the 3 requests/second USDA ceiling. When USDA asks for a longer pause (via Retry-After), that guidance is surfaced in the error text.
• Handle USDA rate limits responsibly: use narrow filters, reuse previous results, and avoid large bulk queries unless necessary.
• Missing or invalid API keys cause the server to log the issue and exit immediately so MCP clients can surface the error.
• Nutrient helpers follow a strict escalation path (scoped abridged → format=full → unfiltered abridged → labelNutrients) and recognise alternate labels such as Total fat (NLEA) so oils and other sparse entries still return macro values. When a Foundation record still hides any macro after those retries, get_macros intentionally errors and points you to alternate FDC IDs or the calorie conversion formula so downstream automations do not ingest partial data unknowingly.
• Legacy SR Legacy identifiers that USDA has retired (currently 4053 for olive oil) are automatically mapped to their documented replacement IDs, and every substitution is called out in the tool summaries (plus requestedFdcIds in the preview payload) so downstream automations can update their catalogs.

Handling Missing USDA Equivalents

Some pantry staples still lack a generic SR Legacy or Foundation record. To avoid thrashing the USDA API:

1. Search for the closest USDA record and document the FDC ID you plan to use (even if it is branded). Ground sumac does have a viable entry—use FDC 2630657 (branded) and record that choice in usda_rebuild_progress.md so other contributors know it is the canonical pick until USDA publishes a generic version.
2. When no acceptable record exists (for example, composite spice mixes that include salt), log the ingredient in usda_rebuild_progress.md under “No USDA equivalent” with the reasoning and the manual macro source. That file now serves as the go/no-go list before you spend time retrying lookups.
3. If you later discover a matching USDA record, update the log with the new ID and remove the “no equivalent” flag so future rebuilds revisit the ingredient.

Development Workflow

• npm run start – Launch with tsx for local development.
• npm run lint – Type-check the TypeScript sources (tsc --noEmit).
• npm run build – Emit the compiled bundle to dist/ for the CLI or packaging.

Run npm run build whenever you change server code and want Codex or other clients that call usda-mcp to pick up the new build.

Troubleshooting

• Startup fails: missing API key – Ensure USDA_API_KEY is exported in your shell or supplied through the MCP client config.
• Client connects but requests fail with 401/403 – Verify the key is active and not rate-limited on the USDA side.
• Repeated 429 Too Many Requests – Each tool already slows calls to one at a time; if you still get 429 responses, wait for the Retry-After duration shown in the error text or batch IDs into fewer round-trips.
• Legacy FDC ID returns “not found” – USDA periodically retires SR Legacy entries. Known substitutions (currently 4053 → 748608 for olive oil) are applied automatically and annotated in the tool response; otherwise, use search-foods to locate the modern FDC ID.
• Foundation macros come back empty – get_macros escalates through full/unfiltered calls plus labelNutrients, including Foundation display labels such as Energy (kcal) and Total fat (NLEA). When a record still hides macros after that sweep, the tool errors with the missing fields. Pick another FDC ID that exposes macros (SR Legacy or Survey datasets usually do) or estimate calories via (fat_g * 9) + (protein_g * 4) + (carbs_g * 4) (oils generally have protein/carbs = 0). FDC 748608 is a known USDA data gap: only the fat numbers are published, so manual macros are required.
• Micronutrients missing – get_micros reports which vitamins/minerals are unavailable in summary.notes. That usually means the USDA record never published the nutrient; check get-food for the raw payload or substitute a richer FDC ID if you need the value.
• No generic entry for an ingredient (e.g., ground sumac) – Document the decision in usda_rebuild_progress.md. Ground sumac uses branded FDC 2630657 as the accepted entry today; for other ingredients without an acceptable record, note the manual macros and rationale so contributors avoid duplicate searches.
• CLI warns about missing dist/server.js – Run npm run build before invoking npx usda-mcp.

For API reference, see the official FoodData Central guide: https://fdc.nal.usda.gov/api-guide.html