Trainerflow Docs trainerflow.com
The app › Connecting AI agents

Getting started

Introduction Concepts Your first program

The language

Values Collections Variables Imports Metadata Feedback Time and rest Grammar

Core elements

Action Input Output Exec Group

Commands

Specify Measure Ask If Repeat While

Examples

Starting Strength Madcow 5×5 An assessment

Standard library

Overview Built-in functions Common Athlete Strength elements Squat Push Pull Bend Core Olympic

The app

Programs and the SPDK The simulator Insights Running a program The mobile app Run notes Trainers and trainees The AI assistant Connecting AI agents Your account

The app

Connecting AI agents

Let your own AI agent — Claude, Claude Code, Codex or any MCP client — work on your programs and runs.

Trainerflow speaks MCP (the Model Context Protocol), so an AI agent you already use can work on your programs the way you would: read them, write them, check and simulate them, and start or stop runs. The agent runs on your own AI plan and spends no Trainerflow credits.

The address

Every agent connects to the same address:

https://trainerflow.com/mcp

There is no API key to copy. The first time an agent connects, it sends you to Trainerflow to sign in and approve it, and Trainerflow issues it access of its own.

Setting it up

The connect page on trainerflow.com has the same steps, with the address ready to copy.

Claude

In Claude on the web, desktop or mobile:

  1. Go to Settings, then Connectors, then Add custom connector, and paste the address.
  2. Claude sends you to Trainerflow to sign in and shows what it is asking for. Approve it.
  3. The Trainerflow tools appear in your chats.

Claude Code

Add the server, then sign in with /mcp inside Claude Code:

claude mcp add --transport http trainerflow https://trainerflow.com/mcp

Codex

One command. Codex spots the sign-in and opens your browser:

codex mcp add trainerflow --url https://trainerflow.com/mcp

If the browser step is interrupted, codex mcp login trainerflow picks it up again.

Any MCP client

Anything that speaks MCP over HTTP can connect. Add a remote server with the address above. If the client asks for a transport, pick Streamable HTTP, and leave any API-key field empty: the client is sent to Trainerflow to sign in.

Approving an agent

When an agent connects, Trainerflow shows you who is asking and what for, and you Approve or Deny. What an agent may do is decided by these scopes:

Scope What it allows
programs:read See your programs, libraries and runs, and read their files.
programs:write Create programs and edit their files and insights. This includes changing a running run.
runs:write Start and stop programs for you and your trainees.

Apps that run on your own computer, such as Claude Code, come with an extra note on the approval screen: “This app runs on your computer.”

What an agent can do

Once connected, an agent can:

  • list your programs, the standard library, your runs and your trainees,
  • create a program and read, write, move, delete and search its files,
  • validate a program and simulate it forward, as prescribed or with the athlete falling short,
  • create, update and delete insights,
  • see your current run and the halt it is waiting on,
  • activate a program for you or a trainee, stop a run, and change a running run,
  • read the Coach language reference, so it writes programs the way Trainerflow expects.

And what it cannot:

  • Train for you. Continuing or stepping back a run stays with the athlete doing it. An agent can see where a run is, but not move it.
  • Send invitations, change the standard library or touch billing.

Each account can make up to 600 agent calls an hour, of which up to 30 can be simulations.

Taking it back

Connected apps, in the user menu, lists every agent you have approved, what it is allowed to do and when it was last used. Press Revoke to cut one off. It loses access immediately; nothing it created is removed, and you can connect it again whenever you like. Deleting your account revokes everything with it.