Skip to main content

MCP Integration

Saturday provides a native Model Context Protocol (MCP) server, allowing AI agents to discover and use Saturday’s nutrition intelligence tools without custom API integration code.

What is MCP?

MCP is an open standard that lets AI models discover and use external tools. Instead of writing custom API client code, an AI agent connects to Saturday’s MCP server and automatically discovers available nutrition tools — their inputs, outputs, descriptions, and safety constraints. Think of it as USB for AI tools: plug in and it works.

Why use Saturday via MCP?

MCP is ideal when your platform uses AI agents that need to dynamically decide when to call Saturday’s tools based on conversation context.

Connecting to Saturday’s MCP server

There are two ways to reach Saturday’s MCP server, depending on who you are:
End users connect via the Claude connector, not an API key. If you’re an individual athlete or coach connecting your own Saturday account to Claude, see the Claude Connector guide — you paste https://api.saturday.fit/mcp into Claude and sign in; no API key needed. The configuration below is for partners integrating their platform with a partner API key.

Server configuration (partners)

Add Saturday to your MCP client configuration with your partner API key:
The server speaks the MCP Streamable HTTP transport and protocol revision 2025-11-25 (negotiating down to older revisions a client offers).

Tool catalog

When your agent connects, it discovers tools including: This is an illustrative slice; an athlete connection currently exposes ~20 tools and a coach connection more (see the coach note below). Each tool includes rich descriptions, structured input schemas, and tool annotations (readOnlyHint, destructiveHint, idempotentHint) that help AI agents use them correctly.

Bottle Builder (interactive MCP App)

build_bottling_plan is an MCP App — alongside its structured result it returns an interactive HTML view (resource ui://saturday/bottle-builder) that renders inline in supporting clients (e.g. Claude.ai). The athlete can drag a strategy slider (even / balanced / concentrated), move fuel between bottles, adjust fill levels, edit their real vessels, and watch carbs/sodium/scoop amounts recompute live. When the target won’t fit the bottles on hand, it returns an honest “carry it more concentrated, top up with water” plan rather than a dead end. Clients that don’t render MCP App views still receive the full plan as text and structured content.
Coach tools. When a coach (Business tier or higher) connects via the Claude connector, an additional set of roster + configuration tools lights up — get_roster, get_roster_digest, get_athlete_fueling_rollup, get_athlete_report, get_session_detail, set_notification_rules, apply_alert_preset, set_report_settings, register_webhook, and more. These are invisible to athlete-only users and to other partners. See the Claude Connector guide for the full coach tool catalog and the Coach API for the equivalent REST surface.

Example: Claude agent with Saturday MCP

Here’s how a Claude-powered agent might use Saturday’s tools in a conversation: Athlete asks: “I have a 3-hour bike race on Saturday. It’s going to be 30C and humid. What should I eat?” Agent’s tool calls:
  1. calculate_nutrition with {activity_type: "bike", duration_min: 180, intensity_level: 8, is_race: true, thermal_stress_level: 8}
  2. search_products with {category: "gel", caffeine: false} (based on athlete preferences)
Agent synthesizes: Uses Saturday’s prescription + product results + safety warnings to give a complete race-day fueling plan.

Safety-aware tool descriptions

Saturday’s MCP tools include safety metadata in their descriptions. This ensures AI agents know that:
  • Prescriptions are guidance for human consideration, not automated commands
  • Safety warnings must be surfaced to the user
  • The not_instructions: true field means “present this to the human, don’t execute it”
Example tool description (what the agent sees):

Tool call example

AI agent guidelines

When building AI agents that consume Saturday via MCP:

Do

  • Surface all safety warnings to the human user
  • Present prescriptions as recommendations, not commands
  • Include “Powered by Saturday” attribution for teaser-tier responses
  • Cache results when inputs haven’t changed — prescriptions are deterministic
  • Handle errors gracefully — if a tool call fails, explain why to the user

Don’t

  • Don’t autonomously act on prescriptions (e.g., auto-ordering supplements)
  • Don’t strip safety metadata from results before presenting to users
  • Don’t modify prescription numbers based on your own logic
  • Don’t use response data for ML training — this violates Saturday’s data policy
  • Don’t make excessive tool calls — batch when possible

LLM discoverability

Saturday publishes machine-readable context files for AI agents: These files follow the llms.txt standard and provide structured context that helps AI agents understand Saturday’s capabilities without reading the full documentation site.