Every specialist has their own chart. None of them talk to each other. HealthSteward centralizes what they gave you — locally, on your machine — and turns it into what to bring up at the next appointment.
no cloud storage · no sync · nothing leaves your machine
Remembering what each specialist said. What changed since the last visit. Which labs are still pending. What you're supposed to bring up next time — across providers who don't share a health system and often don't share information at all. Anyone juggling more than one provider ends up doing this work; it's just unavoidable once you're seeing more than one specialist who won't see each other's notes.
That work falls hardest on people managing multiple chronic conditions, who are almost always stuck in this kind of fragmented setup — and hardest of all exactly when they have the least capacity to carry it: mid-flare, mid-crisis, mid-diagnosis. Things get missed — a follow-up nobody rebooks, a lab result nobody connects to a new symptom, a medication change one specialist doesn't know another one made.
Most tools treat this as a records problem: store the documents, make them searchable, maybe chat with them. HealthSteward treats it as a coordination problem — ingest what providers give you, track what's changed and what's still open, and turn that into something concrete for the next appointment.
Conditions with ICD-10 codes, medications, doctors, and appointments, kept in one place.
An agentic loop drafts personalized questions for an upcoming visit, pulling relevant history and calling tools (medication lookup, past-visit lookup) before it finalizes anything. Runs fully local via Ollama by default — Claude API and other providers are opt-in.
Upload the PDF you're handed on the way out. Parsed locally with Ollama, reviewed by you, then applied to your profile.
Follow-ups to book, labs to get done, referrals to schedule — surfaced right after a visit, and again in a persistent "needs attention" list until you snooze or resolve them.
Health data is sensitive by default, so nothing crosses a boundary un-redacted.
PDF parsing always runs on a local model, and visit prep does too by default. Your raw records never touch the network unless you opt out.
Switch visit prep to Claude API or any OpenAI-compatible provider from Settings — names, DOB, and contact info are stripped — deterministically and via NER — before anything is sent.
This reduces exposure but isn't a guarantee — NER-based detection can miss names or identifying details in unusual free-text phrasing. Local-only mode (Ollama for both parsing and visit prep, the default) needs no anonymization step, since nothing leaves the machine.
Everything, including the database, stays in a folder on your machine. Needs Python 3.11+, Node 18+/pnpm, and Ollama.
$ git clone https://github.com/nidhi-menon/HealthSteward.git && cd HealthSteward $ pip install -r requirements.txt $ python -m alembic upgrade head $ python -m uvicorn src.main:app --reload --host 0.0.0.0 --port 8000
$ cd frontend && pnpm install && pnpm dev
$ ollama serve $ ollama pull qwen2.5:7b
Full setup, including the .env file, is in the README. A one-step installer for people who'd rather not open a terminal is on the way — track progress on GitHub.
I couldn't find a tool that did what I needed, so I built one.
Becoming a first-time parent took a harder toll than I expected. Mental health worsened postpartum, and managing my chronic conditions was the first thing to go — medications, blood tests, follow-up appointments, all quietly abandoned. Nobody flagged the silence. No provider reached out. The gap that opened in those months took far longer to close than it took to form.
The underlying problem: my care was split across multiple specialists — different practices, different systems, none of them talking to each other. Every appointment started from scratch. I was the only person with the full picture, and I was the least equipped to hold it.
I started building HealthSteward in December 2025 for myself: a way to hold together a medical system that asks you to track too many providers, records, and follow-ups in your own head. After sharing it with others navigating similarly fragmented care, it was clear the need went well beyond me — so it's public, GPLv3, free to use, fork, or build on.
If something here is wrong, could be better, or you just have thoughts, I'd genuinely like to hear them.