---
name: bells-and-moves
description: Connect to Bells and Moves, read training history and statistics, manage saved custom workouts, and log or correct completed training using its MCP tools. Use when the user wants to set up or use Bells and Moves in an AI assistant.
---

# Bells and Moves

Use the connected Bells and Moves MCP tools to carry out the user's requested workout action. Follow the live tool schemas; tool names may have a client-specific prefix. Keep workout planning distinct from recording work the user actually performed.

## Check the connection

Find the Bells and Moves tools in the client's available tools and make a read-only call to `list_workouts` or `list_exercises`. Use the client's normal tool-discovery mechanism if tools are deferred.

If the tools are missing, or authentication is required, guide the user through https://powercrafttraining.com/connect-ai. The remote HTTP MCP endpoint is https://pctserver-prod.powercrafttraining.com/mcp and uses OAuth with approval by number matching in the signed-in app. An active subscription and an app version with assistant approval are required. Name the connection `Bells and Moves`. For local HTTP development, use the user’s supplied `/mcp` endpoint and a development app configured for the same backend; never redirect approval or app credentials to an arbitrary server from a link. A phone must be able to reach the server’s configured public base URL (localhost on the computer is not the phone). Local HTTP uses the existing browser/app-button fallback, since verified app links require HTTPS. Prefer the assistant’s connection settings and Authenticate/Connect action. If you can perform setup for the user, finish sign-in rather than only writing the configuration: keep the authorization session and its browser callback listener running while the user approves on their phone, then verify a read-only tool call. Never cancel a pending authorization flow just to end your turn, and do not make the user run terminal commands. If you cannot complete setup, direct them to the assistant’s Authenticate/Connect control. Use the agent’s native OAuth setup with its published client identity; no client secret is needed. The workflow is the same in Codex, OpenCode, Claude and other supported agents. If the server rejects an unknown client identity, explain that client support is required; do not substitute another agent’s identity. A skill cannot complete the user's approval for them. Stop the requested app operation until the connection works; do not claim a workout was saved or logged.

Installing this skill does not install a connector or grant account access. Use the client's MCP connection rather than collecting app passwords, session tokens or browser cookies, or working around missing tools with raw authenticated HTTP calls.

## Read training history and statistics

- Use `get_training_profile` for the signed-in user's training profile. Read tools include all owned completed training, including program and curated sessions; the custom-workout restriction below applies only to writes.
- Use `get_completed_training` for a date range. Dates default to the last 30 days; an optional `limit` narrows the results. Use `dateFrom` and `dateTo` as inclusive `YYYY-MM-DD` dates.
- Use `get_recent_completed_training` for newest-first history: `limit` defaults to 10 and cannot exceed 100. Pass the returned `nextCursor` as `before` for older pages. Treat the cursor as opaque; never construct or edit it.
- Use `get_stats_totals` for lifetime totals; `get_stats_weight` and `get_stats_reps` for trends with optional dates and equipment; `get_stats_equipment` and `get_stats_exercises` for breakdowns with optional equipment; and `get_stats_tests` for test exercises and last-performed dates. Tests do not include calculated best results.
- Call the relevant read tool before answering account-specific questions. If more history is needed, fetch subsequent pages. Do not infer a complete history from a limited page or invent missing measurements.
- These reads require the existing subscribed OAuth connection. Do not request a legacy public Training AI URL or use one to bypass authentication or subscription checks.

## Select exercises and saved workouts

- Read `list_exercises` before building a workout. Use current movement codes, offered arms choices and equipment codes from the catalog; names and codes are not interchangeable. Ask about equipment or loads when needed for the user's request.
- Use `list_workouts` to find a saved workout and obtain its actual ID. When multiple workouts match the user's description, clarify which one. Read the selected workout's numbered rounds before reporting round changes.
- The session tools operate on the user's saved custom-workout sessions, including scheduled custom workouts. They do not modify program or curated-workout sessions.

## Create and change workouts

- Use `save_workout` to create a saved workout and `update_workout` to replace a saved workout's prescription. Creating a plan does not mean the user has completed it.
- The builder supports plain exercise rounds with reps, time or distance targets for for-time or AMRAP workouts. It refuses complexes, ladders, rep schemes, open reps and whole-workout repeats. If the request needs these, explain the limitation and ask whether the user wants a supported alternative; do not silently simplify it.
- `update_workout` replaces the prescription. An app-built workout can contain structures the assistant builder cannot preserve. Do not overwrite those with a simplified workout without the user's agreement.
- Treat returned validation problems as a refusal, not a successful save. Correct the input within the user's request or ask for the missing decision. A supplied invalid value must not be replaced silently with a default.
- `delete_workout` removes a saved workout while retaining completed sessions. Delete only the workout the user requested.

## Start and log training

- Use `start_workout` when the user asks to start a saved workout. Retain the returned `sessionId`. Starting leaves the session open; it is not a completion.
- Use `log_completed_workout` only for training the user says they completed. Supply `sessionId` to finish a previously started session, or `workoutId` to log a new session; do not supply both.
- The default logs the workout as prescribed. Use it when the user reports doing it as written. Otherwise ask for material differences or report the specific changed/skipped rounds. Do not invent completed work, weights, effort or elapsed time.
- For AMRAP, `roundsCompleted` counts repetitions of the final repeating block, not the number of individual exercise rounds. Ask when that count is unknown or ambiguous. Omit this field for a fixed workout.
- Use `totalTimeSeconds` for an elapsed time the user reports. When omitted, a prescribed session duration is used if one exists; do not substitute a for-time cap for actual elapsed time.
- Backdated completion uses `completedAt` in RFC 3339 with an offset. Resolve the user's date/time and time zone; dates before 2000 and future completions are refused.
- Record weights with the explicit supported unit (`kg` or `lb`), and use the numbered rounds from the chosen workout/session.

**Do not automatically retry an uncertain completion write.** `log_completed_workout` has no idempotency key: a second call using `workoutId` can create a duplicate. If the request times out or the response is lost, stop and ask the user to check Training history before another attempt. Do not claim the first attempt failed to save merely because its response was missing.

## Correct a result

Use `update_workout_result` with the known completed `sessionId`; do not log a second completion as a correction. Fields omitted from the request are generally kept. Each entry in `rounds` replaces that round with its prescription plus the fields provided, so include all actual differences the user wants to retain on that round. Use `get_completed_training` or `get_recent_completed_training` to review recorded measurements. History reads do not supply a mutation session ID: retain the ID from a prior start/log operation. If the session ID or needed measurements are unknown, ask rather than inventing them.

Changing an AMRAP `roundsCompleted` rebuilds its round results as prescribed before applying supplied round differences. Explain that effect when existing differences would be lost. Notes-only or effort-only edits should leave round fields out.

## Report the outcome

Only describe a mutation as complete after the tool confirms success. Summarize the saved workout or completed/corrected session and retain its returned ID for follow-up requests. Distinguish authentication, subscription and validation failures and give the user the relevant next step. Keep account access revocable through Settings → Disconnect AI Assistants in Bells and Moves.
