Training

FullFran2stars

A simple Go and SQLite MCP server for conversational workout tracking with ChatGPT

1# Training

README.mdMCP server

Transport: stdio
Status: Local

Power your agent

Add this server

Install from the publisher. Not a PlugMCP endpoint.

Run on PlugMCP

Host a copy with our keys and logs.

No packaged release in the registry yet. Follow the repository README.

RepositoryPublisher installStep-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Server facts

Local

Transport: stdio
License: MIT
Last checked: 2026-10-28

Stars: 2
Tools: —
Uptime: —

Tools

No tool list yet. Remote servers get one after the first health check.

FAQ

Questions

How do I install Training?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does Training need an API key?

No required environment variable is listed on this record.

What transport does Training use?

Training runs locally over stdio.

When was Training last checked?

Last checked 2026-10-28. A simple Go and SQLite MCP server for conversational workout tracking with ChatGPT

From the publisher

Documentation

Training MCP

Log your training by talking to it. A single-user Go service that replaces a workout spreadsheet with 32 conversational MCP tools over Streamable HTTP, plus an installable PWA for the gym floor. SQLite for storage, no framework, no mock library.

you:  "tres por ocho en press banca a 80, RPE 8"
→ log_set  → today's session, created if needed, SI recalculated

Why it is built this way

The same workout gets logged two ways — talking to ChatGPT on the sofa, tapping a phone between sets — and both must apply identical rules. Same validation, same exercise normalization, same SI arithmetic. So the rules live in one domain package and both interfaces are adapters over it:

flowchart LR
    MCP["mcpsrv<br/>32 MCP tools"] --> SVC
    WEB["websrv<br/>PWA"] --> SVC
    SVC["training.Service<br/>rules · validation · SI"] --> PORT(["Store · Clock"])
    PORT -.-> SQL["sqlitestore<br/>SQLite + migrations"]

training imports no adapter. That is what makes it testable without a database, and what makes the two front ends provably unable to disagree about what a set is worth.

Full write-up: docs/architecture.md.

What is worth a look


Ports and adapters, honestly applied	The dependency arrow only points inwards. architecture
`Clock` as a function type	One method, so `time.Now` satisfies it directly. No mock framework in the repo — tests pass a closure.
Prescription copying	Starting a session from a plan copies it. Editing a template never rewrites history; adjusting today never edits the template.
Totals recalculated, never incremented	A deleted or edited set cannot leave a session total wrong.
Advisory, never automatic	`suggest_load` and `volume_recommendation` explain themselves and decline when they lack evidence. Neither ever edits a plan.
Config absence removes surface	Unset `WEB_BASE_PATH` means the PWA is not mounted at all, rather than served on a default path.

Quick path

Install Go 1.26 or newer.
Set MCP_BEARER_TOKEN and run go run ./cmd/training-mcp (default :8080).
Expose /mcp through HTTPS for a public ChatGPT connector; local tunnel-only development is documented in docs/connection-paths.md.
Verify with go test ./... and go vet ./.... CI runs the same checks on every push, plus -race, gofmt, a Docker build, and a 70% statement coverage floor on internal/sqlitestore — the layer that touches the data.

Configuration

Variable/flag	Default	Purpose
`--addr`	`:8080`	HTTP listen address
`TRAINING_DB_PATH`	`~/.local/share/training-mcp/training.db`	SQLite file
`MCP_BEARER_TOKEN`	required	Static bearer credential
`MCP_AUTH_DISABLED=1`	disabled	Unsafe tunnel-only development bypass; never use in production
`WEB_BASE_PATH`	empty (web UI disabled)	Secret prefix the PWA is served from, e.g. `/a1b2c3…`

The unauthenticated GET /health endpoint returns {"status":"ok"}. MCP is only available at /mcp for GET, POST, and DELETE. Public ChatGPT access requires HTTPS at the TLS termination boundary.

Tools

32 tools, grouped by what you are doing: logging sets, managing plans, adjusting today's session, progression and analysis, the feedback loop, and the exercise catalogue.

Full reference: docs/tools.md — every tool with the description the model actually receives, extracted from the source so it cannot drift, followed by the design notes explaining why the surface looks like this.

A few shapes worth knowing before reading it:

log_set needs no session id. It finds or creates today's session, so conversational logging is one call. The explicit start_session + add_set path remains for when you want control.
Load is never planned. A prescription is sets, reps and RPE; the weight that meets it is discovered at the gym.
A set can carry an intensity technique (drop set, rest-pause, myo-reps). It is a property of the set, not a separate exercise, so volume still counts toward the right muscle group while a technique set never becomes the exercise's record.
Every exercise maps to exactly one muscle group, so group SI is a true partition of session SI. Unmapped exercises are reported under an empty group rather than dropped, keeping catalogue gaps visible.

Web UI (PWA)

Setting WEB_BASE_PATH mounts an installable PWA for logging sets from a phone without going through ChatGPT. It is a driving adapter over the same training.Service, so validation, exercise normalization and SI calculation are shared with the MCP tools and cannot drift apart.

Served only under the configured secret prefix; unset means not served at all.
A plan can be started for the day; its exercises become a checklist showing done/target sets, and tapping one loads its prescription into the form.
The day's plan is fully adjustable mid-session: add or drop an exercise, change its set count, skip it, or substitute it when a machine is taken. Sets already logged are never touched by any of this — an exercise removed from the plan simply reappears as off-plan work.
Sets are grouped by exercise, the way a workout is performed and read back.
The entry form shows what was done last time for that exercise and the standing estimated-1RM record; a set matching the record is badged PR.
The exercise field autocompletes from every exercise already known, which is what stops the catalogue fragmenting into near-duplicate names.
Quick-pick chips restore the last weight, reps and RPE per exercise.
Session feedback asks only about the muscle groups actually trained that day.
Superset members are marked in the checklist; after logging one, the next movement of the round loads immediately and the rest timer is held back until the round is finished.
An exercise's setup note is shown whenever that exercise is selected.
/plans builds and edits routines from the phone: rename, add, reorder and remove exercises, and write the routine's notes. It is a separate screen because planning is deliberate work, not something done between sets.
A rest timer starts on every logged set and remembers the adjusted duration.
Sets can be edited in place instead of deleted and re-entered.
Today's session is created lazily on the first set, so opening the app never leaves an empty session behind.
The service worker caches the shell and the last viewed pages for offline reading. Writes always require the network — a set only counts once stored.

GET {base}/export.db streams a consistent snapshot of the whole database, taken with VACUUM INTO so it is safe while the server is running. The result is a plain SQLite file: restoring is copying it back into the data volume. Nothing else in this deployment keeps a second copy, so this is the backup. scripts/backup.sh downloads one, verifies it opens and passes an integrity check, and keeps the last 30 daily copies. scripts/install-backup-cron.sh schedules it daily — a backup you have to remember to take is not a backup.

Authentication is currently the secret URL itself. The reverse proxy injects the bearer token, so anyone holding the URL has full access. This is tracked debt, not a design: see the open issues.

Scope boundary

This MVP excludes recommendations, multi-user identity/RBAC, REST/GraphQL, OAuth/JWT/token rotation, audit history, cloud sync/backups, concurrent-writer controls, and formula changes.

Three exclusions were deliberately lifted as the tool proved useful: frontend/mobile clients (the PWA), analytics (per-muscle volume and exercise progression), and planned workouts (plans). Everything else still stands.

Plans record prescriptions you decide. The volume-landmark feedback loop is implemented as an advisory: it reports a set-count change per muscle group, but never edits a plan on its own.

Documentation

Architecture — ports and adapters, package layout, testing
MCP tool reference — all 32 tools
ChatGPT setup
Connection paths and tunnel safety
SI formula

The external Secure MCP Tunnel smoke flow is unverified in this repository. Use the official provider documentation rather than invented commands.