AI Fitness Coaching

"AI fitness coaching" in 2026 is a confused category — half the products are algorithmic recommendation engines with light branding, the other half are legitimate LLM-driven coaches. The single best-recommended algorithmic-AI app for a decade is still Fitbod; LLM-flavored coaches are improving fast in 2024-26 but are mostly a wrapper over GPT-4 / Claude / Gemini.

Sister sections: Strength Logging, Strength Programming Sources, Training Plans, Self-Hosted AI & LLMs, Apple Watch, HRV & Heart Rate.

Algorithmic-recommendation "AI" (the established players)

• ★ ★ Fitbod — paid (~$13/mo or $80/yr) + 3-workout free trial; algorithmic exercise selection based on your muscle recovery + history; the most-recommended "AI" workout app for 7+ years; clean UI; Apple Health / Health Connect sync.
  • https://fitbod.me/
• ★ Caliber — paid + free; AI + human-coach hybrid; programming + macro tracking.
• Future — paid (~$200/mo); human-coach via app with Apple Watch; not really AI but often listed; expensive.
• MacroFactor — paid (~$70/yr); macro / nutrition; "adaptive" (algorithmic, not LLM) calorie-targeting.
• Aaptiv — paid + free; audio-guided; some AI-flavored class generation.
• Trainerize — paid; a coach app (not consumer); some clients label themselves "AI-powered."

LLM-flavored coaches (the 2024-26 wave)

• Strong AI (Strong) — paid; LLM-overlay on Strong's data; in-app coaching prompts.
• Fitbod AI — added LLM-summary in 2024-25; on top of the algorithmic core.
• Caliber AI Coach — paid; LLM-flavored layer.
• Hevy AI — paid; rolling out 2025-26.
• Whoop Coach — included with sub; LLM-flavored.
• Pulse / Tempo Move AI — niche; hardware-attached.
• Ladder, BetterMe — paid; LLM-flavored content.

Reality: most are GPT-4 or Claude / Gemini wrappers with workout-history context. Quality varies enormously. None replace a real coach for periodization or injury rehab.

Pure LLM as a fitness coach (DIY)

• ★ ChatGPT / Claude / Gemini — free + paid; given a prompt with your weekly volume + last week's lifts + goals, modern LLMs produce surprisingly reasonable beginner / intermediate plans. Not periodization-aware in the way a real coach is.
• ★ Self-host Ollama with a fitness-coach system prompt — totally private; local; quality below frontier models but improving with Llama 3.x / Qwen2.5 / DeepSeek V3. See Self-Hosted AI & LLMs.
• Open WebUI + a fitness coach prompt + your workout CSV uploads — the homelab hybrid. See Open WebUI.

A useful prompt structure:

• "You are an experienced strength coach following [methodology]."
• "Here is my last 8 weeks of training: [CSV / paste]."
• "Goals: [X]."
• "Constraints: [Y]."
• "Generate next week's session structure."

Run / endurance "AI"

• ★ Garmin Coach — free with Garmin watch; adaptive 5K/10K/half/full plans (Greg McMillan etc.). Not LLM, but algorithmic.
• Stryd Power Center — paid + free; power-based plan generation.
• TrainerRoad's AI FTP Detection / AI Plan Builder — paid; well-respected algorithmic plans.
• Final Surge — paid + free; coach-first + some algorithmic features.
• 80/20 Endurance — paid; Matt Fitzgerald's polarized-training plans.

Yoga / mobility "AI"

• ★ Down Dog — algorithmic class generator (not LLM); the longest-running "AI yoga"; see Yoga & Mobility.

Reality-check pitfalls

• No LLM-only coach can replace a clinician for injury rehab, post-surgery return, or pre-existing conditions.
• AI plan generators tend to over-program volume — they don't see your fatigue without explicit input.
• Periodization (deload weeks, peak phases for a meet, taper) is something most LLM-only coaches forget unless you remind them.
• Apple Watch / Garmin's "AI Readiness" is HRV + sleep + RHR + activity; useful but coarse.
• Coach apps will eventually hallucinate exercises that don't exist or recommend exercises that contradict your stated injuries. Verify against Stronger by Science / Renaissance Periodization.

Self-host pattern (private, free)

1. Wger for workout logging. See Strength Logging.
2. Ollama + Llama 3.3 / Qwen2.5 as the LLM. See Self-Hosted AI & LLMs.
3. Open WebUI for the chat front-end.
4. Weekly: paste last week's logs + this week's plan request.
5. Behind Tailscale so it's private.

Cost / license honesty

• Fitbod — ~$80/yr.
• Caliber — ~$30/mo for AI + human hybrid.
• Future — ~$200/mo (premium tier).
• Strong / Hevy / Whoop AI — bundled with their respective subs.
• Garmin Coach — free with Garmin.
• Ollama / Open WebUI — FOSS, free.
• ChatGPT / Claude / Gemini free tiers — usable for fitness-prompt iteration.

Pick this if…

• Algorithmic exercise selection, no LLM: Fitbod.
• Endurance, free with hardware: Garmin Coach.
• Want a real human coach via app: Future.
• Privacy-first, self-host: Ollama + Open WebUI + Wger.
• Frontier-LLM coach: ChatGPT / Claude with structured prompts and your CSV exports.
• Don't trust AI for programming: see Strength Programming Sources for free human-written plans.