Strava AI Running Coach: An Agent AI-Based Personalized Marathon Training Plan Generation System

The Strava AI Running Coach project aims to use agent AI technology to solve the personalization challenges in marathon training. By building a workflow with Claude 4.6 (agent reasoning engine) and the Strava MCP protocol (data access), it generates structured, personalized marathon training plans based on users' actual exercise data, breaking through the limitations of traditional one-size-fits-all training models.

Marathon training requires balancing scientific elements such as load and recovery, and intensity ratio. However, traditional training plans mostly use fixed templates, which are difficult to adapt to individual differences (e.g., physical fitness foundation, time schedule, recovery ability). While manual coaching is personalized, it is costly and hard to scale. Agent AI combines large model reasoning capabilities with personal exercise data analysis, providing a new path to address this challenge.

The core project architecture consists of two main components:

1. Claude 4.6: Equipped with strong reasoning, planning, and tool-calling capabilities, it can understand training principles, analyze user data, formulate plans, and adjust them through multi-round dialogues;
2. Strava MCP protocol: Securely accesses users' exercise data (mileage, pace, heart rate, etc.) from the Strava platform, providing a data foundation for personalized plans.

The agent workflow is divided into four stages:

1. Data Acquisition and Analysis: Obtain historical Strava data via the MCP protocol, analyze training patterns (weekly mileage, pace distribution, etc.) and performance trends;
2. Goal and Constraint Identification: Understand users' marathon goals (finishing/PB), time constraints, venue conditions, and preferences;
3. Plan Generation: Divide training cycles (base/intensification/peak/taper phases) based on analysis results, and formulate weekly and single-session training content;
4. Feedback and Adjustment: Dynamically optimize the plan based on user execution feedback (e.g., adjust intensity or recovery arrangements).

Technical implementation needs to focus on three key points:

• Data Privacy: Use a token mechanism to ensure user data security, and users can revoke access rights at any time;
• Scientific Validation: Set up safety check mechanisms to ensure progressive training load, reasonable intensity distribution, and sufficient recovery;
• User Experience: Present plans in a clear and easy-to-understand way, provide execution guidance and encouragement to improve user compliance.

The project has a wide range of application scenarios:

• Beginner runners: Provide complete guidance from entry to finishing;
• Advanced runners: Help break through bottlenecks and prepare for PB;
• Time-constrained users: Maximize training efficiency. Its methodology can be extended to sports such as triathlon and cycling, and even to areas like nutrition planning and rehabilitation training.

Future development directions include:

1. Multimodal Data Fusion: Integrate data such as sleep, nutrition, and stress to build a more comprehensive health profile;
2. Real-Time Adjustment Capability: Dynamically adjust training content based on the day's physical state;
3. Community Learning Mechanism: Learn from a large number of user training results to optimize recommendation algorithms.