PredictFit connector documentation

The PredictFit connector lets the AI assistant you already use (Claude or ChatGPT) read the parts of your PredictFit record you allow and add entries you mention. This page is the reference for what it does. For step-by-step setup see Connect your AI.

What it is

PredictFit is an iPhone app that keeps the Apple Health data you choose - weight, nutrition, workouts, sleep and resting heart rate - in one record and projects it forward. PredictFit does not run AI itself. The connector is a standard Model Context Protocol (MCP) server. Your assistant calls it to read your record and to queue entries for your app to save.

• Reads are served from the last sync of your own phone. The app computes scores and projections; the connector only serves them, and every read says when the data was last synced.
• Writes never go straight into your record. The assistant queues an entry; your PredictFit app saves it on its next sync and labels it "Added by AI". You can edit or delete any of them in the app.
• Requires the PredictFit iPhone app with an active subscription (one plan, $4.99 per month, first month free, billed by Apple).

Connector address: https://predictfit.app/mcp

Authentication

The connector uses OAuth 2.1 (authorization code with PKCE) and dynamic client registration, so Claude and ChatGPT can connect without any key or password being pasted. When you add the connector, your assistant opens a PredictFit sign-in page:

1. Sign in with the same PredictFit account you use in the app (Apple, Google or email).
2. Choose which data areas the assistant may read (see "Data areas" below) and whether it may add entries.
3. Approve. Your assistant receives a token that works only for your account and only for what you ticked.

Access tokens are short-lived and refreshed automatically. Removing the connector in your assistant, disconnecting in the app, or deleting your account ends access immediately. An inactive subscription also pauses the connector until it is restored.

Tools

Each tool carries MCP annotations so your assistant knows how to treat it. Read tools are marked readOnlyHint: true and idempotentHint: true. Write tools are marked readOnlyHint: false and destructiveHint: false: they only add, never delete or overwrite, and you can reverse them in the app. No tool reaches outside your own PredictFit record (openWorldHint: false).

Read tools

• get_coaching_context (Read coaching context) - one call that returns your profile, active goal, coaching state, streaks, trends, recent life context, constraints and a plain-language summary.
• get_health_summary (Read health summary) - overall Health Loop score, domain scores, readiness state and top drivers.
• get_projections (Read projections) - weight projection, strength forecast and readiness forecast computed by the app.
• get_opportunities (Read opportunities) - a ranked list of coaching opportunities from the decision engine.
• get_goals (Read goals) - your active goals plus profile basics (age, height, weight, sex).
• get_weekly_summary (Read weekly summary) - the weekly narrative your app shows you.
• get_logging_schema (Read logging schema) - the field shapes the write tools accept and the categories already in use, so the assistant reuses them.

Every read includes a freshness stamp (when your app last synced). If the app has not synced for 36 hours the read is flagged stale and the assistant is told to say so. If an area has never synced, the tool explains why and what to do instead of guessing.

Write tools

• propose_food_log (Add food log) - a meal with calories and macros, from a description or a photo, with the assistant's confidence.
• propose_body_metric (Add body metric) - body weight, waist or body fat measurement, or a blood pressure reading.
• propose_observation (Add observation) - a qualitative observation: a 0-10 scale (energy, stress, soreness, mood...), a behavior, life context, a symptom, a journal note, a decision or a milestone.
• propose_behavior_log (Add behavior log) - a behavior mentioned in passing, for example alcohol or a missed walk.
• propose_symptom_log (Add symptom log) - a symptom, for example bloating or a headache.
• propose_activity_log (Add activity log) - activity your wearable did not capture, for example moving firewood for an hour.
• propose_daily_reflection (Add daily reflection) - a daily reflection or note.
• propose_goal (Add goal) - a new goal you agreed to in conversation (weight loss, weight gain, maintain, strength, endurance, general health). An existing goal of the same type is updated rather than duplicated.
• propose_goal_update (Add goal update) - a change to a goal target or to daily calorie and protein targets.

Every write requires the assistant's confidence (0.0 to 1.0) and stores what you actually said. The assistant is instructed never to invent values and to ask when unsure. Writes are idempotent when the assistant supplies an external id, so a repeated message does not create a duplicate.

Data areas (scopes)

At sign-in you choose any of five areas. Reads are filtered to the areas you granted; an area you did not grant is not served at all.

• Sleep - duration, quality and sleep trends.
• Cardio - steps, heart rate and endurance activity.
• Nutrition - meals, calories, macros and weight trend.
• Strength - workouts and strength progress.
• Recovery - readiness, rest and recovery signals.

Adding entries is a separate permission ("read + add entries"). Without it the assistant can read what you granted but cannot queue anything.

Limits

• At most 20 AI-added entries per 24 hours per account. The 21st attempt is refused and the assistant is told the entry was not saved.
• Entries are validated: unknown categories, out-of-range scale values, implausible calories or weights, and malformed dates are rejected with a specific error rather than saved.
• Reads have no fixed quota; they return the app's last published values, so calling them again does not change anything.
• The connector cannot delete entries, change your Apple Health permissions, change your subscription or delete your account.

Privacy summary

• Only the areas you granted are served, only to the assistant you connected, only while the connection is active.
• Each tool call is recorded in an activity log you can see in the app (tool name, time, client), with parameters stored as a hash, not as text.
• PredictFit does not read your conversation. The connector receives only the arguments of each tool call.
• What Claude or ChatGPT does with data it reads is governed by Anthropic's or OpenAI's own terms and privacy policy.
• Deleting your PredictFit account removes your record, queued entries, connections and activity log from PredictFit's servers and invalidates every token. See Delete your account.

Full policy: Privacy policy. Terms: Terms of use.

How to disconnect

• From the app: open PredictFit, go to your profile, and remove the connected assistant. Access stops immediately.
• From your assistant: remove the PredictFit connector in Claude or ChatGPT settings.
• To change what the assistant may read or add, disconnect and connect again, choosing new permissions at sign-in.

Support

Email support@predictfit.app with the assistant you are using and any error text shown. See also Support.