The engine behind your coach.

Most fitness AIs are a model and a prompt. Fitly is a ten-layer orchestrator: nine systems run before the model is allowed to speak, and a tenth runs after. Here is what happens beneath the chat bubble.

0
orchestration layers per turn
0+
structured tools
0
context queries in parallel
0
output guards on every reply

The model is the last thing we trust.

The naive AI app takes your message, glues on some profile data, sends it to a model, and renders the reply. It works for demos and collapses the moment a real person shows up with a real life, a shoulder injury, a hatred of quinoa, and a "make it harder" with no other context.

A model is a smart, expensive, occasionally-wrong text engine. Everything that makes Fitly feel like a coach instead of a chatbot is engineered around it.

Ten layers run on every message.

You type one sentence. Nine systems run before the model replies, and a tenth runs after.

1

Auth, validation & limits

Every request is verified, schema-checked, and rate-limited before a single token is spent. The layer that says no is what lets the rest feel generous.

2

Image classification

Photos are sorted into eight types (meal, label, barcode, scale, equipment, body, screenshot, other) so each one triggers the right pipeline instead of one fuzzy guess.

3

Parallel context assembly

Nine queries fan out at once: profile, macro targets, today's logs, recent workouts, active plans, history, workflow state, long-term memory, and streaks.

4

Workflow continuation

Detects whether "yes, the bigger one" is a fresh request or a reply mid-task, across nine multi-turn workflows, so the conversation just flows.

5

Dynamic prompt builder

The system prompt is assembled per turn from 20+ conditional blocks: personality, diet, targets, tools, refusal rules. Every irrelevant token is a tax on quality.

6

Dynamic tool scoping

50+ tools exist; the model only sees the few that fit this turn. Smaller tool lists mean faster, more accurate tool choices.

7

Multi-model routing

Cheap tasks run on a fast model; hard ones route to a stronger one, and a turn can escalate itself mid-flight. Thoughtful coaching at consumer prices.

8

Agentic tool loop

The model can call tools, read results, and call more, up to ten iterations, before it replies. One bubble can hide six tool calls. It checks its work.

9

Response guards

Three guards run before the reply ships: it really saved what it claimed, it can't write into future days, and it can't quietly drop a logged food.

10

Persistence & memory

After the reply, the turn is stored, memory candidates are extracted, streaks advance. The layer that makes the next conversation sharper than the last.

Want the full deep dive?

We wrote a long-form breakdown of all ten layers, what runs, why it runs, and what it costs to get right.

Read: The Orchestrator Inside Fitly →

An agent other agents can talk to.

Fitly isn't just an agent you talk to. It runs a Model Context Protocol (MCP) server, so your Claude, ChatGPT, Gemini, or any custom agent can connect and use the same tools your coach uses. Your fitness data becomes a capability any agent can call, no app required. Available on the Pro plan.

Your AI agents
ClaudeChatGPTGeminiCustom
→MCP
Fitly AI
your agent + 50+ tools
→
Your fitness data
read & write
Read your profile, goals, and history
Log meals, workouts, weight, and water
Check macros and today's progress
Pull up and activate your plans
Scoped, token-based access per agent
Streamable HTTP transport

What the orchestrator buys you.

An agent, not a prompt

The model is one component, not the product. It reasons, picks from 50+ structured tools, calls them in a loop, and only answers once it has checked the data. That is the difference between a chatbot and an agent.

Full context, every turn

Before the model speaks, nine queries assemble your present-tense state: goals, injuries, allergies, today's macros, recent workouts, and long-term memory. It never asks you to repeat yourself.

Provider-agnostic, multi-model

Fitly runs on frontier models (Claude by default) behind an abstraction, and routes each turn to the right one for the job, escalating mid-turn when a task turns out to be hard.

Photo intelligence

A fast classifier sorts every image into eight categories first, so a nutrition label gets OCR'd, a barcode gets looked up, and a plate gets portion-estimated, each by the right pipeline.

Guards you can trust

A write-confirmation guard, a future-write guard, and a partial-logging guard run on every response. We don't let the model lie about what it saved, log into next Tuesday, or drop a food.

Private by design

Health data is encrypted, served through pre-signed URLs, and every data change is written to an audit log. Your data is yours; we never sell it.

Sharper every week

After every turn, Fitly extracts what it learned about you into long-term memory, so the next conversation starts further ahead than the last.

Built to scale

A stateless API, cursor-based pagination, Redis-cached context, and queue-backed processing on AWS, architected to serve millions from day one, not bolted on later.

Everyone has the same models.
The difference is what runs around them.

Every fitness AI in the App Store has access to roughly the same models we do. The orchestration around the model is the moat, and it is where we have spent our build cycles. That is why a sentence as ordinary as "I had a chicken bowl from Chipotle" can do as much real work as it does inside Fitly.

Download on the App Storeor open Fitly on the web →

Join the waitlist.

Drop your email and we'll let you know when spots open up.