Three files, three names for the same test, one LOINC code. The agent finds all three, aggregates the trend, and names the file every number came from.

What Mirobody does

• One record for the whole family. Invite a partner, a parent, even a child who never signs in at all, and keep the household's health history in one place.
• Every source, one record. Garmin, Oura and Whoop connect directly; anything already written into Apple Health comes with it; PDFs, phone photos, spreadsheets, exports: 23 file types in all, and Mirobody reads them.
• No hallucinations, everything traceable. Every indicator lands in one settled system: either it gets a definite code, or it says it could not resolve one. It never invents one in between. Built and tested against real reports, in English, Chinese and Japanese.
• The agent reasons only over coded data. Trends by minute, hour, day, week or month, drawn as a chart; a baseline and how far a number has moved in one call; comparisons across labs, files and devices, because one standard (LOINC and UCUM) sits under all of them. It reads medications and genetic variants too.
• Runs on a laptop. Four containers, 791 MiB resident, under 5% CPU idle. No GPU, no Node.js.
• Your model, your key, your data. Model calls go to the model you chose. Everything else stays on your machine.

Try it in 60 seconds

One command, five spellings: watch which ones it recognises, and which one it refuses. No key, no config, no network, and with uvx, no install either:

uvx mirobody resolve "LDL cholesterol" 血红蛋白 ヘモグロビン "空腹血糖(GLU)" 血脂

血红蛋白 and ヘモグロビン: two languages, one code, 718-7. 血脂 (lipids) names a category, not one observation, so it resolves to nothing. The resolver would rather return nothing than guess a code, because a wrong one puts two different tests on the same trend line.

from mirobody.engine import resolve, resolve_reading

resolve("血红蛋白").loinc                                # '718-7'   any language, one code
resolve("total cholesterol").loinc                     # '2093-3'  [Mass/volume]
resolve_reading("total cholesterol", "5.0", "mmol/L")   # '14647-2' [Moles/volume]
resolve_reading("total cholesterol", "193", "mg/dL")    # '2093-3'  the unit picks the code

resolve("中性粒细胞百分比").loinc                          # '26511-6' Neutrophils/Leukocytes
resolve_reading("中性粒细胞", "62 %", None).loinc          # '26511-6' a percentage...
resolve_reading("中性粒细胞", "4.2", "10*9/L").loinc       # '26499-4' ...and a count are two codes
resolve("血脂").resolved                                 # False    a category, not an observation

Pass the value and the unit when you have them. A different unit means a different test, and LOINC folds that into the code's own identity, so one name is deliberately several codes. → Engine reference · Indicators

Collect · Translate · Agent

An indicator takes three steps from arriving to being cited. Each one leaves a trace, so the answer at the end can be followed back to the page it came off:

① records how the source spelled it, ② decides what it actually is, ③ answers on that footing. Comparing a number across two labs, charting three years of it, computing a baseline: all of it rests on the code ② hands over.

The agent does not have to be ours. Every tool it uses is served at /mcp as well, gated per user. Claude Desktop, Cursor or your own loop run the same tools over the same record, and get back the same indicators.

Garmin, Oura and Whoop connect with your own credentials from each vendor; the setup guide walks it through. Apple Health goes another way: a client on the phone hands the data over, so any band, ring or scale reaches your record the moment it writes into Apple Health, with nothing to integrate here at all.

Privacy

Nothing leaves your machine except calls to the model you chose. Reading a photo of a report, pulling indicators out of a PDF, answering your question: all three call it. Which provider and which model is the one key in your .env.

② Translate stays local entirely: a name to a code, a unit to UCUM, looked up against a bundle that ships inside the package. No key, no network, no GPU, no model. Your record lives in your own Postgres, in containers you run, and nothing here reports usage anywhere.

One key, and it is the only secret you hold. Put an OpenRouter key (OPENROUTER_API_KEY), a Gemini key (GOOGLE_API_KEY), an OpenAI key (OPENAI_API_KEY) or an Anthropic key (ANTHROPIC_API_KEY) in the .env beside compose.yaml, then docker compose restart. DeepSeek, DashScope or any OpenAI-compatible gateway works alone too. Which model chats, which reads report photos, which extracts indicators and which embeds are four lines in config.llm.yaml, and that file names the variable (api_key: OPENROUTER_API_KEY), never the secret. mirobody doctor prints what each surface selected, and names the fix where one has nothing.

The quickstart ships its secrets as placeholders, and encryption at rest does not yet cover every field. Before this reaches a network you do not control, read SECURITY.md: it also lists exactly what the server calls off your machine.

🚀 See it end to end

git clone --depth 1 https://github.com/thetahealth/mirobody.git && cd mirobody
git lfs install && git lfs pull   # the resolver's LOINC bundle, 40 MB; a fresh clone holds a pointer stub until you do
./deploy.sh                       # Postgres + pgvector, Redis, server, worker → http://localhost:18060

(--depth 1 skips the history of superseded frontend builds; drop it if you plan to send a pull request.)

Sign in as you@mirobody.ai, code 111111, no mail provider needed. An account of your own is one request away:

curl -X POST localhost:18060/password/register -H 'Content-Type: application/json' \
     -d '{"email":"me@example.com","password":"at-least-8-chars"}'

SEED_DEMO_DATA is on by default, so two accounts are already there with 2,019 indicators between them: you, and mom@mirobody.ai, who shares her record with you view-only. Set it to false to hold real data and neither account is created. Settings → Add member covers someone who will never sign in at all, a parent, a child, with a record you hold on their behalf.

Drop a file on the Data page and watch it become indicators. demo/upload/ holds four files the seed deliberately leaves out: a lab PDF, a phone photo of a printed report, a spreadsheet and another lab's CSV export. Each analyte comes out with a value, a unit and a LOINC code, linked back to the page it was read from.

Ask how the cholesterol has moved and the agent finds every file that carries it: one lab writes Cholesterol, Total where the others write Total Cholesterol-TC, and both are 14647-2. It charts the trend and names the file each number came off: 4.60 → 4.45 → 4.38 mmol/L. Ask for a baseline or a monthly average instead and the same tool aggregates over the whole record, rather than handing back rows for the model to add up itself.

Ask the same question of the record shared with you and it is a different person's answer, from data you can only view. Sharing is invite-only, off by default, and strictly permission-checked.

→ The four-minute walkthrough · examples/06_care_circle_rules.py prints the whole sharing decision table offline · Docker deployment · Configuration

Check any of it yourself

Every figure below comes with its source: a command you can run, or a public dataset.

• 213/213 on the tests an ordinary checkup prints, in English, Chinese (Simplified and Traditional) and Japanese. The set is deliberately the least flattering one — everyday panels, written the way a report prints them, which is what every new user tries in their first minute. test_engine_coverage.py prints the score when you run it.
• Three open benchmarks, public datasets, one command each: longitudinal health agents, medical hallucination, harmful medical advice. mirobody-eval · datasets · arXiv:2604.02834.
• The package names the vocabulary that answered you: mirobody.BUNDLE_VERSION → loinc-2.82+2026.08.28-af2524b7a285, the release, the cut date, and a digest over the bundle's own contents.
• 305 standard pulse indicators and 326 UCUM units with dimensional analysis. The full counts, and why the bundle holds at LOINC 2.82 rather than 2.83, are in Standardization in depth.
• pip install mirobody is 2 packages, numpy the only dependency.

The engine powers Theta Wellness, a live consumer health product with 5,000+ registered users.

🔌 Use it, extend it

→ API overview · MCP integration · Adding tools · Bringing your own agent

🤝 Contributing

The highest-leverage contribution is a term the resolver gets wrong. Run mirobody resolve "<term>"; if the answer is wrong or empty, report it or add a row to resolver_overrides.tsv plus a case to test_engine_coverage.py — the coverage score is the review.

pip install -e '.[test]' && pytest -q && lint-imports

→ CONTRIBUTING.md · Local Python setup · Repository layout · Roadmap · CHANGELOG · SECURITY

📚 Documentation, and what shaped this

docs.mirobody.ai, in English and Chinese — start at the Quickstart or the API reference; contributor guides are in docs/.

Mirobody's design draws on the following standards and projects, with thanks: HL7 FHIR, Regenstrief Institute (LOINC), UCUM, OHDSI OMOP, Open Wearables, Open mHealth / IEEE 1752, wearipedia, dlt / Airbyte / Singer, deepagents and LangChain. The terminology licences this ships under are in LICENSE-3RD-PARTY.

If it read a report for you, a star helps the next person find it. Releases land most weeks — Watch for them.

Apache 2.0 · © 2026 Theta Health