Ironstack

Train with a partner. Own your data. See everything.

Page: missionbuilt.io/rack/ironstack. Ironstack lived inside the Loadout repo until September 2026; it has its own repo now, with its history.

Ironstack turns Claude into your training partner and Elasticsearch + Kibana into your training log. You log sessions conversationally — Claude asks the questions a good partner asks, writes the log, and keeps you honest and motivated. Your data lands in your own Elasticsearch, and seven Kibana dashboards read it back: what you lifted, how hard it was, where you were, and how it felt.

Then you can ask it things:

"What did I squat a couple weeks ago?" "How did I feel when I was traveling in Las Vegas?" "What cue fixed my bench setup last block?"

Like the skills in The Loadout, the skills here are plain markdown — no installer, no API key, no telemetry. And like Floodlight, Ironstack keeps your data where it belongs: every rep you log lives in your own private repo and your own Elasticsearch. Nothing here phones home.

Not medical advice. Ironstack logs your training and encourages you. It does not build training programs, and it does not assess pain or injuries — it will point you to qualified professionals instead. Nothing in this project is a substitute for a doctor, physical therapist, or qualified coach.

What's here

ironstack/
├── skills/          ← the Claude skills (drop into .claude/skills/)
│   ├── workout-partner/    log + motivate: your training partner with a notebook
│   └── workout-recall/     ask your training history anything
├── coach/           ← the Kibana agent behind ASK THE COACH: prompt, tools, privileges
├── starter/         ← the plumbing: copy this into your own PRIVATE repo
│   ├── schema/             JSON Schemas (workout, meet) + the index mappings
│   ├── ingest/             log.py, setup_indices.py, index_workouts.py, index_meets.py
│   ├── workflows/          GitHub Action → place at .github/workflows/index.yml
│   ├── config/             defaults, the exercise taxonomy, the equipment registry
│   ├── workouts/           where your logs go
│   └── meets/              where your meet records go (two public examples included)
├── examples/        ← one fictionalized session (.iron + .json + .md) showing the format
├── kibana/          ← build_dashboards.py → dashboards.ndjson, import.py, the dashboard map
└── CEILING.md       ← the load ceiling, written once for every surface that suggests a weight

Skills are shared and public. They live here, in this repo. Your data is private. It lives in a repo you create from starter/, plus your own Elasticsearch. The two meet in your Claude session.

How it works

you + Claude (workout-partner skill)
        │  conversation during or after your session
        ▼
workouts/2026/2026-09-01.iron            ← written once, by hand
        │  python ingest/log.py … --push
        ▼
… .json + .md in your private repo       ← generated, never hand-edited
        │  git push → GitHub Action validates & indexes
        ▼
your Elasticsearch, seven indices
        │
        ├─ Kibana: Overview → Program / Session / Lift / History / Meets / Mindset
        ├─ the Ironstack Coach, if you build one (coach/)
        └─ you + Claude (workout-recall skill): ask your history anything

The seven indices

Index	One doc per
workout-sessions	session, including digest — the whole day written out as a paragraph
workout-sets	set, with the derived numbers on it: e1RM, intensity, INOL, muscles, equipment
workout-notes	note, watch items included
workout-meets	competition attempt
workout-daily	training day, pre-aggregated
workout-weekly	ISO week, pre-aggregated: ACWR, monotony, strain, projected total
ironstack-signals	verdict row — the data behind the dashboards' Signal cards

ironstack-signals carries no date-typed field on purpose. Every row is windowed when it is written, so the time picker cannot quietly re-scope a verdict after the fact.

Get started

1. Stand up Elasticsearch

Pick one:

• Elastic Cloud Serverless — create an Elasticsearch project (docs). ELSER semantic search over your notes works out of the box.
• Free self-hosted — curl -fsSL https://elastic.co/start-local | sh, or your own Docker setup. The core app runs on the free tier; the semantic layer switches on wherever it's supported (ES_SEMANTIC=auto).

Create an API key with write access and note your Elasticsearch endpoint URL.

2. Make your private instance repo

Copy starter/ into a new private repo (see starter/README.md for the commands), and move workflows/index.yml to .github/workflows/index.yml. Then add two Actions secrets:

• ES_ENDPOINT — your Elasticsearch URL
• ES_API_KEY — the write API key

Every push that touches workouts/** or meets/ now validates and indexes automatically. No CI? Run python ingest/setup_indices.py, python ingest/index_meets.py, and python ingest/index_workouts.py by hand — all three are idempotent.

3. Import the dashboards

export KIBANA_URL=https://<your-project>.kb.<region>.gcp.elastic.cloud
export ES_API_KEY=<key with Kibana saved-object privileges>
python kibana/import.py

(Or Kibana → Stack Management → Saved Objects → Import → kibana/dashboards.ndjson.) Switch Kibana to dark mode; the palette assumes it.

Seven dashboards, one navigation graph, every aggregate a door to its detail:

Dashboard	What it answers
Overview	Where the block stands. Three Signal verdicts, what you wrote about it in your own words, days to meet, the projected total against your meet best, and the block timeline as the door into a session
Program	Block, week, day. How hard this week is loading, in words; every day in the range; the weekly loading table
Session	One session, start to finish: header, top set, tiles, every set with its warm-up, notes in order, wrap-up, conditions, PREV / NEXT
Lift	One exercise over time: where it sits against its own best, e1RM with that best drawn on it, where the reps land by zone, every working set
History	How heavy this block is against earlier runs of the same kind, reps by zone, the session timeline, acute vs chronic load, the sessions table
Meets	The platform record: whether this run-in matches the last one, totals, DOTS, best lifts, every attempt
Mindset	Everything you wrote: what you keep writing down, the tag chart, recent notes, the sessions behind them

Click a bar, a point or a row and you land on its detail with the context carried as a filter: Overview → block timeline → Session → prev / next. kibana/README.md has the full map, the drilldown table, and what Lens can and cannot do here.

Three values are set at build time, not in the saved objects:

Variable	What it does
IRONSTACK_COACH_URL	Your coach's URL. Set it and every dashboard gets an ASK THE COACH panel. Unset, no panel is built — and --no-coach is how the committed dashboards.ndjson is built, so nothing here points at anyone's deployment
IRONSTACK_MEET_MAX_LB	Your best competition total, in pounds. Draws the reference line on Overview's projected-total chart, which is a Lens chart and cannot read your data. The card beside it reads the real number out of workout-meets either way
IRONSTACK_TZ	A fixed offset (UTC, -07:00) used to format dates and compute days-to-meet

Rebuilding is python kibana/build_dashboards.py with those set, then python kibana/import.py. With no coach of your own, build with --no-coach; an unset IRONSTACK_COACH_URL is an error rather than a silent removal of the link from all seven dashboards.

4. Train with Claude

Add the two skills from skills/ to Claude (Claude Code, Cowork, or claude.ai skills):

• workout-partner — log sessions conversationally. It writes one .iron file in your private repo and runs the one command that generates the rest.
• workout-recall — connect Claude to your Elasticsearch (e.g. the Elasticsearch MCP server) with a read-only API key and ask your history anything.

Each skill folder carries its own README, licence and references, so copying the folder copies everything it needs. See examples/ for what a logged session looks like.

5. The coach, if you want one

The ASK THE COACH button on the dashboards points at an agent you build in Kibana Agent Builder: read-only tools over your seven indices, the same load ceiling the skills obey, and the same two refusals. coach/ has the system prompt, the tools with the ES|QL each one runs, and the minimum privileges for its API key.

Agent Builder availability varies by Elastic tier and version. Everything else works without it — the logs, the indices, the dashboards, and both skills are independent of the coach.

Conventions

• Effort is one scale: RPE (10 = max). Reps-in-reserve converts as RPE = 10 − RIR.
• Every session carries its program context: block, phase (hypertrophy / strength / peaking), week, day of N, and the meet date. The indexer adds prev / next session links, streak day, and days to meet; never write those by hand.
• The .iron shorthand is the only thing written by hand. The JSON and the markdown are generated, so they cannot drift apart.
• Meets are logged in kg (the source of truth); pounds are derived.
• Carries log as walks with distance_ft; holds as seconds; bodyweight work as weight_lb: 0.
• Location is coarse by design — town or city only. Traveling? Log the city and travel: true, and "how did I feel in Vegas?" becomes answerable.
• Mindset notes carry tags from a small taxonomy (felt-strong, body-awareness:<area>, grip, cue, …) so themes trend over time.
• An exercise name that isn't in config/exercises.json is an error, not a warning. A lift that silently becomes a new name splits its own history in two.

Design principles

1. Your data is yours. Public skills, private data. Your repo is the source of truth; Elasticsearch is a rebuildable projection of it.
2. Claude is the interface. Capture conversationally, recall conversationally. Dashboards are the glanceable layer.
3. One rule, one place. The load ceiling is in CEILING.md and every surface that suggests a weight cites it. Two copies of a rule are two rules.
4. AI-first, gracefully optional. Semantic search over your notes (ELSER semantic_text) is designed in from day one, and everything degrades cleanly to structured and full-text search where it isn't available.
5. No cost barrier. Runs on Elastic Cloud Serverless or a free self-hosted Elastic.
6. Partner, not programmer or doctor. Honest encouragement and in-the-moment lift cues, yes. Programs and injury advice, never.

License

MIT, like the Loadout skills. Fork it, learn from it, make it yours.

---

On The Rack · Mission Built. Real strength is lifting others.