PlatePal is an MCP server. Keep chatting with your assistant and snapping photos of your meals. PlatePal remembers everything in structured form, does the math, and gives your assistant the context and coaching habits to notice what a good trainer would.
https://platepal.xyz/mcpThe fix isn't a smarter model. It's structured state the assistant can read, and procedures that make it read that state before giving advice.
Photo or text in, structured data out. Your personal food library learns the things you eat every day ("the usual hummus lunch"), and estimates stay honest: restaurant sodium is logged as a range, not a fake-precise number.
Before any advice, the assistant gets a briefing: training streak, load against your norm, when each muscle was last worked, sleep and recovery, symptoms, and the coach's notes. It also remembers your personal form cues ("your back rounds on RDLs past 155 lb").
A dense day-by-day timeline across food, training, sleep, pain and weight, plus precomputed signals for what's moving. The assistant can notice "sodium is creeping up" or "you're getting stronger" without being asked.
Structured Postgres with row-level security, so each person's data is walled off at the database level. Export and deletion are first-class features.
The assistant stays the whole interface. PlatePal does the deterministic work (totals, unit conversion, training load, trends) and supplies the coaching procedures, so improving the coaching never requires reinstalling anything. It's agent-agnostic: anything that speaks MCP can connect.
"Let's do legs tomorrow."
A coach with a briefing sees you've trained seven days straight, run 18 miles this week, your HRV is below baseline and your back was tight after Saturday's squats. So it suggests an easy day tomorrow and low-spinal-load legs the day after, and says why.