AI agent

You already run your coaching from the dashboard. AI agent lets you run a lot of it by chat instead — from the AI assistant you already use. Connect your Claude or ChatGPT to Protocol, then just ask: “Build a 4-week deload for my powerbuilding group,” “Who hasn’t logged in 10 days?,” “Summarise Mike’s last check-in and draft a reply.” Your assistant does the work in Protocol — directly and live — and shows you the result.

The key idea: you bring the AI, Protocol brings the tools. Protocol doesn’t run the model and doesn’t charge you for it — your assistant (your Claude or ChatGPT account) does the thinking, and Protocol gives it a safe, scoped set of actions on your account. Think of it as a second way into the same Protocol you already use, not a new place your data lives.

What it can do

Three things, and it’s genuinely good at all three:

• Build & edit programs — describe what you want and it builds it live: a reusable library template by default, or assigned to a client directly when you say so (“build a 4-week PPL and assign it to Dejan”). Bulk edits too: “swap barbell bench for dumbbell across weeks 2–4.” → Building & editing programs
• Analyse your roster & read reports — open-ended questions your dashboard’s fixed views never anticipated: adherence trends, who’s gone quiet, this-quarter-vs-last comparisons, a summary of any client’s latest check-in. → Analysing & summarising
• Triage your day — “What needs my attention this morning?” — a plain-language rundown across your whole roster, from your phone, no dashboard open.

What it can’t do (by design)

• It can’t see more than you can. The agent is scoped to your account only — the same clients, Vault, and reports you already have access to. It can’t reach another coach’s data.
• You set the access, and you can disconnect anytime. When you connect you approve how much access to grant — read only, read + write, or read + write + send — and that level is the boundary on everything it does. The Claude plugin works the same way: its consent screen asks you to pick a level too, with read + write preselected. See Installing the Claude plugin for what its consent screen actually grants. Disconnect it anytime and it stops immediately. See Safety, scopes & privacy.
• Hard limits no agent key can cross. An agent key is the connection almost every coach uses (the AI Agent page, the MCP connection). No agent key, at any scope, can message your clients (it can read conversations but never send a client a message), touch billing (charges, refunds, subscriptions), permanently delete anything, or generate with Protocol’s own AI (your assistant is the AI). These are enforced by the platform: there’s no tool that does them, at any scope.
• One credential where that isn’t true: a REST API key. A pk_live_ key, created under Integrations, API keys (owner only: only the account owner can see that screen and create one), authenticates as you on every route, the same as signing in yourself. None of the four limits above are blocked at the platform level for that key; they hold only because the assistant chooses to follow them as policy. It’s reachable only from Claude Code, and only if you set it up deliberately. Almost no coach needs it. See Safety, scopes & privacy and Installing the Claude plugin.
• It’s only as good as the AI you connect. You choose the assistant; a stronger model gives a better experience. Protocol gives every assistant the same clear tools, but it can’t make a weak one smart.

Basic vs. advanced

Pick a guide: