Skip to content
← Back to projects
Case study

Pichurri Fitness

"Pichurri" is what I call my AI agent cluster. It comes from a colloquial, affectionate Spanish word — something like calling someone "little one" or "buddy", but warmer and harder to translate. Think of it as a pet name for software that knows too much about you.

An AI agent that works as a real personal coach: programs, remembers, insists, analyzes and adapts. Not a chatbot — a system.

24/7 coach via Telegram20+ automated cron jobsPersistent memoryProgression tracking
What it is

An AI coach that doesn't wait for you to write.

Pichurri Fitness is an autonomous agent running 24/7 via Telegram. It designs workouts, tracks nutrition, manages supplementation and maintains an accountability system. Part of a multi-agent ecosystem with 5 specialized agents (dev, trading, media, fitness, coordination), each with an exclusive domain.

Architecture

OpenClaw + Claude + Telegram + persistent memory.

  • Multi-agent platform OpenClaw running on NAS 24/7
  • Base model: Claude (Anthropic) for reasoning and analysis
  • Channel: real-time bidirectional Telegram
  • 20+ scheduled cron jobs (nutrition, training, supplements, hydration, check-ins)
  • Heartbeat system that detects silence and writes proactively
  • Ability to delegate tasks to other agents in the ecosystem
Memory

Every session builds on the previous one.

  • Daily files with exercises, weights, sets, nutrition and metrics
  • Long-term memory with goals, patterns and decisions
  • Semantic search across all history before responding
  • Per-exercise progression tracking with full history
  • Pattern detection (skipped exercises, changing adherence)
Real-time coaching

You train and report. The agent responds instantly.

During training, I send results as I finish exercises. The agent suggests weights based on history, records actual data, and if equipment is busy, reorders the session or suggests alternatives. Exercise changes are backed by current research.

Running

I send screenshots and it takes care of the rest.

After each run, I send Apple Watch screenshots via Telegram. The agent extracts the data (distance, pace, heart rate, splits) and compares with previous sessions to see if there's real progression.

Accountability

"No Mercy" mode 💀

Every night, full check-in: meals, training, supplements, water, reflection. The agent detects patterns and doesn't accept vague excuses. If I say "I'll try to go", it responds "No trying. Tomorrow Upper A."

Result

It's helping me progress consistently.

The combination of proactive cron jobs, persistent memory, real-time data analysis and a direct coaching tone creates an experience that goes beyond a fitness chatbot. The daily structure and constant tracking make it much easier to stay disciplined.

Stack
OpenClawClaudeTelegramNode.jsCronApple Watch