Skip to main content
Glama

rocket

An adaptive marathon training planner for one runner.

The primary interface is conversation: Claude reads and writes plan state through MCP tools. A thin PWA renders the calendar and a 30-second daily check-in. The system plans an aggressive but guarded block, absorbs real-world disruption, and re-plans continuously from Garmin metrics plus subjective feedback.

Non-negotiable invariants

  1. Guardrails beat enthusiasm. Requests that violate ramp-rate, recovery or taper rules are negotiated, never silently executed.

  2. Degrade gracefully. Every feature must work on cached data and manual check-ins alone. Garmin is an upgrade, not a dependency.

  3. The plan is a rolling window. Only 7-10 days of concrete sessions exist at once; replanning must be cheap.

  4. Every actual activity updates load state, however it arrived.

  5. Context tags matter. Surface, elevation and footwear change training stress for the same distance. Never score on distance alone.

Related MCP server: claude-garmin

Where to start

  • docs/specs/ — the source of truth, eight files

  • REDLINES.md — non-negotiables, and what happens when one is violated

  • AGENTS.md — operating rules for anyone, human or agent, working here

  • docs/decisions.md — append-only decision log

  • docs/ci-gates.md — every gate this project should have, built or not

Engineering discipline

Adapted from discipline-kit. Two ideas carry most of the weight:

  • A gate you have not proven fails is a decoration. When a gate is added, it is broken on purpose, watched to catch, reverted, and the violation recorded in its ledger row.

  • A gate you cannot build yet is a NOT IMPLEMENTED row, never a silence.

Every commit carries a Task: tasks/<slug>.md trailer, enforced fails-closed on pre-push. Nothing runs on git commit.

Licence

MIT.

Related MCP Connectors

Related MCP Servers

  • A
    license
    A
    quality
    B
    maintenance
    Enables Claude to access and query your Garmin Connect data, including sleep, activities, training load, and health metrics, through a set of read-only MCP tools.
    28
    109 PyPI
    1
    MIT
  • A
    license
    Not graded
    quality
    B
    maintenance
    MCP server that connects Garmin Connect data to Claude, enabling training analysis, recovery checks, and personalized plans based on real metrics like HRV, training load, and activities.
    17
    MIT
  • F
    license
    Not graded
    quality
    B
    maintenance
    Connects Claude to Garmin Connect data for personalized running coaching, including morning readiness checks, post-run analysis, weekly reviews, and goal tracking.
    -
  • F
    license
    Not graded
    quality
    C
    maintenance
    AI-powered coaching for runners, cyclists, swimmers, and triathletes, enabling personalized workouts, training plan adaptation, performance analytics, and AI coaching via Claude, MCP clients, or HTTP.
    -