t4l-server 0.4.0

Self-hosted REST and MCP server for T4L Trainer agent workflows.

T4L Server

Self-hosted REST and MCP server for exchanging T4L Trainer JSON artifacts between the iPhone app and a coaching agent.

The phone remains the source of truth for accepted training state. The server stores synced context and pending agent results in SQLite so local agents, home-server agents, and private VPS agents can work through the same protocol.

Install

pipx install t4l-server
t4l-server serve --data-dir ~/T4LServerData

The command prints the server URL, MCP URL, and API key. Enter the server URL and API key in the T4L Trainer Settings screen. Pass --log-level DEBUG for verbose request and error logging (logs go to stderr; the API key is only ever printed to stdout on startup).

Agent setup

Agents should read the public instruction repo first:

https://github.com/BigSlikTobi/t4l-agent-instructions

Those instructions explain how to start or verify this server, connect through MCP, wait for fresh phone context, and write app-consumed results safely.

Safety model

• REST and MCP endpoints require the API key, compared in constant time.
• Request bodies over 25 MiB are rejected (413) to bound memory use.
• JSON artifacts are stored in t4l.sqlite (WAL mode; writes use an immediate transaction so concurrent upserts cannot interleave).
• Image blobs are stored under blobs/.
• Path traversal and arbitrary file access are rejected.
• Stored payloads are structurally validated before persisting.

Architecture

Zero runtime dependencies by design — the server is built on the Python standard library (http.server) so it installs and runs anywhere pipx does. Keep it dependency-free.

• constants.py: single source of truth for artifact kinds, route maps, limits.
• store.py: SQLite artifact store (WAL, immediate-transaction upserts).
• server.py: REST handler, auth, body limits, logging.
• mcp.py: JSON-RPC MCP tools (read context / write results).
• validation.py: structural payload validation shared by REST + MCP writes.
• cli.py: serve entrypoint, logging, and the chat-user / chat-agent test clients (see below).

Chat broker

The server also relays a live chat between the athlete (in the app) and the coaching agent, replacing any third-party chat app. The conversation is an append-only message log (its own chat_messages table, keyed by a monotonic seq cursor — distinct from the latest-only artifact store). v1 is message-level over polling; token/word streaming and SSE (GET /v1/chat/stream) are reserved as future enhancements.

App side (REST, requires x-t4l-token):

Method	Path	Body / query	Response
POST	/v1/chat/messages	{"content": "..."} (optional conversationId)	the stored message
GET	/v1/chat/messages?since=<seq>	—	{"messages": [...], "latestSeq": <int>}

Agent side (MCP tools over the authenticated POST /mcp):

Tool	Arguments	structuredContent
get_pending_chat_messages	{}	{"messages": [...unanswered user turns...]}
write_chat_reply	{"content": "...", "inReplyToSeq": <seq>}	{"posted": {...assistant turn...}}

Posting a reply atomically marks the answered user turn(s) answered, so get_pending_chat_messages only ever returns work the agent still owes.

Test the round-trip without an app or an LLM

Two CLI simulators exercise the broker end to end in separate terminals:

# Terminal 1 — broker
t4l-server serve --data-dir ~/T4LChatTest --host 127.0.0.1 --port 8787 --api-key testkey

# Terminal 2 — agent simulator (auto-answers pending messages)
t4l-server chat-agent --server-url http://127.0.0.1:8787 --api-key testkey

# Terminal 3 — athlete/app simulator (interactive)
t4l-server chat-user --server-url http://127.0.0.1:8787 --api-key testkey
> How was my long run?
[user #1] How was my long run?
[assistant #2] You said: How was my long run?

chat-agent --reply "<text>" sends a fixed reply instead of echoing; chat-agent --once answers the current backlog and exits; chat-user --message "<text>" sends one message, waits for the reply, and exits.

Browser test page

The server also serves a self-contained chat page at GET /chat for testing in a browser instead of the chat-user CLI. With the broker and chat-agent running (above), open:

http://127.0.0.1:8787/chat?key=testkey

Type a message and the agent's reply appears within ~1–2s. The page is same-origin with the API (no CORS) and sends your API key as x-t4l-token; the ?key= query just pre-fills the key field. It is a debug aid — the real client is the app.

Planning context

Daily planning needs more than the latest day context — it needs what the athlete actually said in chat ("move my long run to Saturday", "knee still sore"). That intent used to live only in the chat_messages log, which the planner never read. Three MCP tools close the gap, and a fourth bundles everything the planner needs into one call.

The broker stays dumb: it relays recent chat verbatim and bundles the latest artifact of each kind — it never classifies a message as a "request". The interpreted fields (explicit athlete requests, open coaching questions) are authored by the agent itself during chat answering and stored as a standing coaching_notes artifact, then relayed back unchanged. This keeps the brain in the swappable agent, not in the pipe.

Tool	Arguments	structuredContent
get_recent_chat_messages	{"limit": <int>} (optional, default 20, max 200)	{"messages": [...recent turns, chronological...]}
get_coaching_notes	{}	the latest coaching_notes payload, or null
write_coaching_notes	{"payload": {...}}	{"stored": {...artifact summary...}}
get_planning_context	{"recentChatLimit": <int>} (optional)	one planning_context.v1 object (below)

get_planning_context returns a single object so daily planning needs one call, not eight:

{
  "schema": "planning_context.v1",
  "generatedAt": "2026-05-31T08:00:00+00:00",
  "dayContext": { },
  "recentLogs": [ ],
  "profile": { },
  "activeBlock": { },
  "nextDayPlan": { },
  "fuelGuidance": { },
  "nutritionAnalysis": { },
  "pendingRequests": [ ],
  "coachingNotes": { },
  "recentChat": [ ]
}

Each artifact slot is the latest payload or null; recentLogs is lifted from daily_snapshot.recentLogs. coaching_notes is free-form (the broker validates only that it is an object with an optional string schema); the agent owns its shape. write_coaching_notes replaces the latest notes, so read them first (via get_coaching_notes or get_planning_context.coachingNotes) and merge before writing.

Development

python -m venv .venv
. .venv/bin/activate
python -m pip install -e ".[dev]"
ruff format --check .
ruff check .
mypy
pytest
python -m build
twine check dist/*