# I Connected Apple Health to an AI Agent (Hermes) > [!question] Lingering Questions > - The entire hosting/deployment/AI-provider stack is provided or discounted by the video's sponsor (Hostinger/Nexus.ai); alternative self-hosting paths aren't discussed. How portable is this build to a non-sponsored provider? > - Biometric data (sleep stages, HRV, heart rate, body composition) is transmitted to and stored on a third-party VPS with no discussion of encryption at rest or data retention policy. Worth flagging against the vault's local-model-sovereignty material. > - "Self-learning" is asserted but not technically detailed — is Hermes's memory RAG over stored conversation, a structured user profile, or something else? Left unresolved by the source. > [!tip] Article Ideas > - Test whether Ollama + a local LLM can replace the sponsored Nexus.ai provider in this stack to keep biometric inference fully local — a genuine infrastructure question, not just a privacy preference. > [!link] Cross Reference > - [[AI Agent Safety MOC]] — MCP-based self-configuration and third-party biometric data storage both raise trust/consent questions > - [[AI & the Future of Knowledge Work MOC]] — persistent agent memory as the differentiator between a tool and a longitudinal assistant > [!quote] Vocabulary > - *Hermes Agent* — the self-hosted AI agent in this build that maintains memory across sessions and configures its own tool access (e.g., Grafana) via MCP > - *HealthBridge* — the free iOS app (by Alex Morris) that reads Apple Health exports and pushes them directly to a PostgreSQL connection string, handling ingestion without custom code ## Metadata - Source: [[3. Resources/Readwise/Articles/I Connected Apple Health to an AI Agent... Here's What Happened]] - Format: YouTube video, sponsored build walkthrough ## Author's thesis Gardas demonstrates that a meaningful, memory-enabled AI health coach is achievable at hobbyist scale using free/open-source tooling (HealthBridge, PostgreSQL, Grafana) plus a low-cost VPS and a self-hosted agent (Hermes). The differentiating claim is persistent cross-session memory: rather than treating each health measurement as an isolated data point, the agent interprets today's numbers against recent history (illness, previously ignored suggestions), producing advice calibrated to the individual rather than population baselines. The build is presented as a proof-of-concept blurring the line between an infrastructure project and a personal wellness tool, while carrying real but undiscussed privacy tradeoffs — full biometric history stored on a third-party VPS — and a sponsorship bias toward one specific hosting/AI-provider stack. ## Key arguments (≤3) 1. **Persistent memory changes what a health AI can say, not just what it remembers** — knowing the user was sick last Tuesday or has ignored morning-workout suggestions twice changes the interpretation of today's data, not just the historical record. 2. **MCP lets an agent self-provision a tool integration from a natural-language instruction** — telling Hermes about Grafana is enough for it to install the needed MCP and gain working access, with no manual API wiring by the user. 3. **Composable Docker services on a commodity VPS lower the bar for personal AI infrastructure to hobbyist level** — the "Lego block" framing describes stacking PostgreSQL, Grafana, and an agent on one VPS via one-click deploys, not a custom systems-engineering project. ## Claim atoms ### Persistent agent memory reframes AI coaching as longitudinal relationship > [!quote] ID:rw_healthagent_01 > It knows I was sick last Tuesday, so it doesn't penalize the low step count. It knows I don't like morning workouts because I already ignored those suggestions twice now. Where/why: already promoted — see Permanent candidates below. ### MCP enables AI agents to autonomously configure third-party dashboards > [!quote] ID:rw_healthagent_02 > Open your Hermes agent and tell it about Grafana. It will install the MCP needed for the configuration. And that's it. Hermes now has access to your health data. Where/why: a distinct, generalizable claim about agent self-provisioning — not yet captured by an existing Permanent note. Promoted below. ### One-click Docker orchestration lowers self-hosted AI stack complexity to hobbyist level > [!quote] ID:rw_healthagent_03 > You can run multiple Docker apps on a single VPS, snapping them together like Lego blocks until you've got the full solution. Where/why: an aphoristic framing of composable self-hosted infrastructure — noted but not promoted to a standalone Permanent note (thin on its own, mostly restates the architecture description). ## Permanent candidates - **Already captured**: the memory-as-coaching-context claim — [[3. Resources/Permanent Notes/Persistent Memory Turns Health Metrics Into Coaching Context]] (created 2026-08-11, same source). Not duplicated here. - **New**: the MCP self-provisioning claim (atom H2) is distinct and not yet in the vault — see the new Permanent note below. ## Processing checklist - [x] Checked for existing Literature Note — none existed prior to this pass - [x] Checked for existing Permanent notes from this source — found one covering the memory/coaching claim; created one new note for the distinct MCP self-provisioning claim - [ ] Consider a thin Permanent note on the sponsorship-bias/undiscussed-privacy pattern if a second source corroborates it — currently single-source - [x] Connect into the relevant MOCs — [[3. Resources/MOCs/AI Agent Safety MOC]], [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] - [[3. Resources/Permanent Notes/Persistent Memory Turns Health Metrics Into Coaching Context]] — the memory/coaching claim from this same source, already in the vault