clawcoach

Your own strength coach — in the app you already have open.

It writes your programme, texts you when it's time to train, and rebuilds next week around what you actually lifted this week. Miss a session and it adjusts. Hit every rep at RPE 6 and it adds weight. Tell it your shoulder hurts and it swaps the movement — then remembers, permanently.

No app to install. No subscription. No feed. Just a coach that messages you first, built on OpenClaw and running on hardware you own — your training data never leaves it.

⚠️ Not professional advice

clawcoach is a hobby project driven by a language model. It is not a trainer, physiotherapist, or doctor, it is not a medical device, and it is no substitute for professional advice. It can be confidently wrong.

You use it entirely at your own risk and are responsible for your own safety. Consult a physician before starting any exercise programme. If it hurts, stop.

Read DISCLAIMER.md in full before using it — and especially before setting it up for someone else.

you                          coach
─────────────────────────────────────────────────────────
                             12:00  Training day. Day A:
                               DB Floor Press  3x8 @ 15lb  RPE 7
                               DB Row          3x8 @ 20lb  RPE 7
                               Band Pull-Apart 3x15
                               Bike 20 min easy

done, floor press felt
easy at 15

                             Logged. Floor press → 17.5lb Friday.
                             e1RM up 4% in two weeks.

Why this beats asking a chatbot about fitness

A chatbot waits for you. This one comes looking.

• It chases you. It schedules its own check-ins and fires them whether or not you feel like training. That's the entire difference between a programme and a good intention.
• It builds on your actual numbers. Every session is logged. It reads your recent history before prescribing anything, so week 4 is a response to weeks 1–3 — not a fresh guess that happens to look similar.
• It does the maths. Estimated 1RM, volume load, weekly training load — computed in code, not eyeballed. It will tell you your e1RM rose 4% before you'd have noticed.
• It programs like a coach, not a search result. Progressive overload, RPE/RIR-based loading, deloads, WHO/ACSM cardio baselines — and it cites its source when new evidence changes your programme.
• It respects your time. Numbers over adjectives. No pep talks, no "Great question!", no restating what you just said. Most messages are under eight lines.

Guardrails

Not optional extras — they are the reason this is safe to hand to a family member.

• Never programs through pain. Soreness is fine; pain is a stop.
• Never invents a number. If it doesn't have the data, it asks.
• Refers out for anything medical — plainly, one line, no hedging.
• Under 18: body composition is off the table entirely, and pain means stop and tell a parent — not the adult 12-week referral window.
• Never pushes weight loss unless the trainee raises it first.

Prerequisites

This is a coach agent for OpenClaw — it does not install OpenClaw. Before anything here works you need:

1. A running OpenClaw gateway on a machine you control (a spare laptop or small server is plenty). Check with openclaw gateway status — expect running and probe ok.

2. A model API key configured in OpenClaw, e.g. Anthropic. The coach is a normal agent; it uses whatever model your gateway is set up with.

3. The Docker sandbox image built. Channel-originated sessions run sandboxed, and without the image every message fails with "Something went wrong":

   git clone --depth 1 https://github.com/openclaw/openclaw.git ~/openclaw
   sudo apt install -y docker-buildx     # docker.io ships without buildx
   bash ~/openclaw/scripts/sandbox-setup.sh
   docker images | grep openclaw-sandbox

   If you just added yourself to the docker group, reboot — the systemd user manager that spawns the gateway keeps its old group set, and restarting the gateway is not enough.

4. A Telegram account for whoever will use the coach, and the ability to create a bot via @BotFather.

5. SSH access to the host, since setup runs there.

You do not need a public IP or an open port. Keep the gateway on loopback and reach it over a VPN or tunnel — see SECURITY.md.

Quick start

git clone https://github.com/YuvalShalev/clawcoach.git
cd clawcoach
./scripts/setup-agent.sh coach-sam      # run on your OpenClaw host

That creates the workspace, copies the agent files in, and puts it under git so a bad edit is recoverable. It is the only automated step.

The remaining three need your own secrets, so they are deliberately manual — a script taking a bot token as an argument would leave it in shell history:

1. Create the Telegram bot and verify the token before it touches config: curl -s "https://api.telegram.org/bot<TOKEN>/getMe"
2. Get the trainee's numeric Telegram ID from @userinfobot — this is the allowlist, and without it anyone who finds your bot can talk to it.
3. Add the agent to openclaw.json with openclaw config patch, then openclaw gateway restart.

⚠️ agents.list is an array, and arrays replace on patch — read your current list and write it back in full, or you will delete your other agents.

docs/SETUP.md has the exact JSON and the mistakes worth avoiding. Budget about 15 minutes.

Layout

agent/          the coach itself — AGENTS.md is the control surface
  AGENTS.md     method: intake, programming, logging, review
  SOUL.md       personality and hard limits
  PROFILE.md    ─┐
  PLAN.md        │ empty templates; the agent fills these in
  LOG.md         │
  METRICS.md    ─┘
scripts/        setup-agent.sh, push.sh, pull.sh
docs/SETUP.md   installation, tools, scheduling, troubleshooting
SECURITY.md     the tradeoffs, honestly — including the unsolved ones
DISCLAIMER.md   health & safety — read before running it for anyone

Adding a second person

One agent per person — LOG.md holds one training history, so two people cannot share a workspace. Each needs their own bot, workspace, and Telegram ID.

The agent files are identical for everyone. Age is data, not method: intake records it in PROFILE.md and the under-18 rules key off that. No per-person forks to maintain.

Tuning it

agent/AGENTS.md is the control surface. When behaviour is wrong, edit that file and restart the gateway. Expect iterations — most of what is in there now came from watching it get something wrong.

Security

It handles health data and has web access. Read SECURITY.md before running it for someone else — it documents the rw workspace tradeoff and what prompt injection can still do.

For health and safety, and what you take on by setting this up for another person, see DISCLAIMER.md.

Status

Running daily against a real training block. The agent files are stable; the setup scripts are lightly tested outside their original host.

License

MIT