MCP vs web app capabilities

Use MCP for

Connecting an AI assistant via OAuth (recommended) or API key — see Connecting to MCP.
Scripted or assistant-driven pantry, supply, meal plan, and Galley edits.
Semantic search and match_meals from an AI IDE.
Receipt → pantry ingestion when the agent's own LLM has done the parsing (preview_inventory_import → apply_inventory_import).
Bulk meal-plan edits, supply lifecycle (complete_supply_list), and user-preference updates.

MCP does not include

These AI-heavy features exist only in the web app and consume Ration AI credits there:

Receipt scan (Ration runs the OCR/LLM)
AI meal generation
AI plan week
URL recipe import
Vector embedding generation for cargo writes (skipped on MCP for cost reasons)
get_credit_balance (UI-only surface)

There are no MCP tools that replace those flows. The principle is: if the agent's own LLM can do the AI work for free, MCP exposes the deterministic data path. If Ration would have to spend credits, the user must opt in through the web UI.

REST vs MCP

Need	Surface
CSV inventory export/import	REST `inventory` scope
Galley JSON export/import	REST `galley` scope
Supply CSV export	REST `supply` scope
Interactive agent tools	MCP host — OAuth (recommended) or API key with *`mcp:`** scopes (see REST API (v1) overview scope table)

Credits reminder

MCP reads and writes do not debit the organization’s AI credit balance. They still change real data—confirm destructive actions in your assistant.

When product marketing or Pricing lists features, those lists override this summary if they conflict.