I built an AI strength coach because I wanted my training backed by real studies A developer created PerformanceAgent, an open-source AI strength and conditioning coach that runs inside AI agent CLIs like Claude Code. The tool uses a deterministic Python engine for calculations and a verified local evidence corpus to prevent hallucinated citations, aiming to provide realistic, science-backed training advice. πŸ‹οΈ The first open-source AI Strength & Conditioning Coach powered by scientific research. English Β· FranΓ§ais /clementrx/Performance-agent/blob/main/docs/i18n/README.fr.md Β· EspaΓ±ol /clementrx/Performance-agent/blob/main/docs/i18n/README.es.md Β· Deutsch /clementrx/Performance-agent/blob/main/docs/i18n/README.de.md Β· Italiano /clementrx/Performance-agent/blob/main/docs/i18n/README.it.md It runs inside an AI agent CLI β€” a terminal program you chat with, such as Claude Code, Gemini CLI, or Codex β€” and turns it into a professional coach that designs, explains, monitors, and adapts your training. No backend, no API key, no hosting, no extra subscription. And it will tell you the truth when your goal is unrealistic: bash $ claude I run 10K in 55:00. I want 35:00 in 12 weeks. πŸ‹οΈ Coach: I evaluated your goal with the feasibility engine: Probability: 0.2% β€” I have to be honest, this is unrealistic. You're asking for a 36% improvement β€” about 3%/week sustained for 12 weeks. Beginners sustain roughly 1%/week. Counter-proposal: 46:30 in 12 weeks ~78% probability , then we reassess. Want me to build that program? LLM fitness coaches have two failure modes: they invent scientific references, and they tell you what you want to hear. PerformanceAgent is architected so neither is possible: LLMs narrate, the engine calculates. Every number β€” feasibility probabilities, race predictions, training loads, periodization waves β€” comes from a deterministic, property-tested Python engine. The agent explains the math; it never does the math. Citations can't be hallucinated. The coach may only cite studies returned by the local evidence corpus graded, DOI/PMID-verified . The PDF renderer hard-fails on any reference that isn't in the corpus. Your data is files, not a cloud. Profile, programs, session logs, and check-ins live in a plain directory of markdown/YAML you can read, edit, diff, and sync. PerformanceAgent isn't an app you open β€” it plugs into an AI agent CLI. Once plugged in, you just talk to it in plain language; no config files, no commands to memorize. Never used Claude Code before? Install it first: curl -fsSL https://claude.ai/install.sh | bash full instructions: code.claude.com/docs https://code.claude.com/docs/en/quickstart.md . You'll also need uv https://docs.astral.sh/uv/getting-started/installation/ β€” it fetches the right Python version by itself, nothing else to install. Step 1 β€” plug in the coach. Run this once, from any terminal: claude mcp add performance-agent -s user \ --env PERFORMANCE AGENT HOME=~/athlete-data -- uvx performance-agent This registers the coach's "brain" the engine, the science library, your future athlete profile as a tool Claude Code can call. -s user makes it available in any folder you later open claude from. ~/athlete-data is just a suggested path β€” pick any folder, it doesn't need to exist yet: the coach creates it the first time it saves something. That's where all your data lives as plain files; nothing is sent anywhere. Step 2 β€” teach it how to coach. Step 1 gave Claude the tools the math, the data . This step gives it the coaching protocols β€” when to ask what, when to be honest about a goal, how to build a program: git clone --depth 1 https://github.com/clementrx/Performance-agent mkdir -p ~/.claude/skills cp -R Performance-agent/skills/ ~/.claude/skills/ Step 3 β€” fully quit and restart Claude Code. New tools load only when a claude session starts : close any open session completely and run claude again. Check it worked β€” in the fresh session, ask: List the performance-agent tools. You should see 47 tools. If so, you're done β€” just talk to it. Open a terminal and start your agent claude . Say your goal in plain language β€” any language works. "I want to run a 10K under 50 minutes" , "PrΓ©pare-moi pour un Hyrox" . Answer the coach's questions. First time, it runs a short onboarding current level, history, schedule, equipment and saves your profile. Get an honest verdict. The feasibility engine scores your goal; if it's out of reach, you get the real probability and a counter-proposal. Say yes β€” the program is written to athlete/programs/program-v1.md , periodized into cycles with deloads and taper, every prescription carrying its purpose, evidence grade β˜…β˜…β˜…β˜…β˜… β†’ β˜…β˜†β˜†β˜†β˜† , and verified citations. After each workout, log it in one line β€” "Logged: 8 km easy, RPE 4" . Every couple of weeks, say "Check-in" β€” the coach compares plan vs. reality, quotes your training load, and flags drift. When life happens β€” "I tweaked my ankle" , "I can only train twice next week" β€” the program is adapted and versioned program-v2.md with the reason recorded in the audit trail. Feasibility numbers below come from the deterministic engine β€” run it yourself and you'll get the same ones. Want to see the resulting program files? Each of these three conversations ends with a full program β€” browse them in examples/ /clementrx/Performance-agent/blob/main/examples , in all five languages. I want to run a 10K under 50 minutes. I have 5 months. Coach me. πŸ‹οΈ Coach: Two facts before I judge that goal: your current 10K time, and how long you've been training consistently? I ran 55:12 last month. About a year of running, 3Γ— a week. πŸ‹οΈ Coach: Feasibility over your 21-week horizon: Probability: 84% β€” realistic, and worth committing to. You need a 9.5% improvement, i.e. 0.45%/week, while a runner at your training age can sustain about 1%/week. Comfortable margin. Want me to build the program? Yes. πŸ“„ Program written to athlete/programs/program-v1.md 21 weeks, 3 runs/week: one interval session, one tempo, one long run β€” periodized with deloads and a 10-day taper. I signed up for a Hyrox in 3 months. Build me a plan. πŸ‹οΈ Coach: Hyrox is 8 km of running broken up by 8 strength stations β€” so we train both, together. Three facts: your 10K time or best guess , your strength background, and days per week you can train? 10K around 52 min, I lift twice a week, I can do 4 sessions. πŸ‹οΈ Coach: Good base for a strong first finish. Honest note: 12 weeks is enough to finish well and enjoy it; chasing a podium time would need a longer runway β€” I'd rather tell you now. Plan on 4 days: one running-interval day, one station-strength day sled work, lunges, wall balls , one "compromised running" brick run β†’ station β†’ run, the skill Hyrox actually tests , one easy long run. Deloads weeks 4 and 8, race-week taper. Build it? Yes. πŸ“„ Program written to athlete/programs/program-v1.md I want visible abs and bigger arms in 3 months. πŸ‹οΈ Coach: Those are two different physiological goals β€” arms grow with volume and enough calories; abs show at low body fat, which usually means a deficit. Before I promise anything: height, weight, training experience, and sessions per week? 1.78 m, 82 kg, two years of lifting, 4 sessions a week. πŸ‹οΈ Coach: Honest verdict: in 12 weeks you can make real progress on both, but not maximize both β€” recomposition is slow at two years of training. The realistic version: - Moderate calorie deficit with a high protein target the Nutritionist skill quantifies it, with safety guards . - Upper-body priority: 12–16 hard sets/week for arms, heavy compound work kept in to protect strength in the deficit. - 3 lifting days + 1 conditioning day. - Expected at week 12: visibly fuller arms, abs "outlined" if you start around 15–18% body fat; leaner than that takes longer, and I'll say so at every check-in rather than let you guess. Want the program? Yes. πŸ“„ Program written to athlete/programs/program-v1.md The MCP server engine, science, data speaks a standard protocol β€” MCP https://modelcontextprotocol.io β€” that most AI coding tools understand. The coaching skills are packaged in Claude Code's native skill format; elsewhere you paste the same instructions into that tool's own "custom instructions" file. | Client | MCP tools | Coaching skills | |---|---|---| | Claude Code | βœ… native steps above | βœ… native | | Gemini CLI | βœ… native | GEMINI.md | | Codex | βœ… native | AGENTS.md | | Cursor | βœ… native | .cursor/rules/ .mdc | | Windsurf | βœ… native | | | VS Code GitHub Copilot | βœ… native | .github/copilot-instructions.md | | Cline VS Code extension | βœ… native | .clinerules/ | Setup commands for each, PDF reports requires typst , data-directory resolution, and troubleshooting: docs/installing.md /clementrx/Performance-agent/blob/main/docs/installing.md . Any other tool that supports MCP servers works with the same uvx performance-agent command. Just here to use the coach? Skip this β€” it's for the curious and for contributors. php flowchart TB U You <-- H Your agent CLI