mirobody

⚡ Try it in 60 seconds

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

pip install mirobody instead if you want it on your PATH. Either way the resolver answers from a bundle that ships with the package.

血红蛋白 and ヘモグロビン land on the same code as hemoglobin, LOINC 718-7; 空腹血糖(GLU) on fasting glucose. 血脂 (lipids) is a category, not an observation, so the resolver returns nothing rather than a plausible wrong code.

…

Pass the value and the unit when you have them. LOINC encodes the unit and the result type into the identity, so the same name resolves to different codes. resolve abstains rather than guessing: an empty code is a gap worth a second look; method="refused" is a decision. → Engine reference · Indicators

Why Mirobody

• Two reports, same test, different spelling and units. That is the problem. Mirobody translates any reading, in any language, to LOINC + UCUM — and refuses to guess when it does not know the term.
• AI can only reason over what it can read. The agent reads the original documents, charts lab draws against sensor-derived series, and cites the page it read from.
• Self-hosted, standards-based, Apache-2.0. One ./deploy.sh runs the whole stack on your own machine; what comes out is LOINC-coded, FHIR-ready records you can take anywhere, and the same tools are served over MCP to Claude Desktop, Cursor or your own agent. No GPU: the app is self-hosted, the model is not. You bring one API key to a hosted provider; nothing runs inference on your machine. "Offline" here means the resolver — name to code, with no network and no key — not a local LLM.

Collect · Translate · Agent

The engine does three things, and the codebase, the docs and Contributing are organized around exactly these three stages:

By the numbers

Why 2.82 and not 2.83, what LOINC covers of the wearable world, and the opt-in semantic tier that cannot abstain: → Standardization in depth

Benchmarks — open and independently reproducible

The most-downloaded health-AI benchmarks in their category on Hugging Face, 4,000+ downloads a month each: ESL-Bench (event-driven longitudinal health agents — 100 synthetic users, 10,000 queries, arXiv:2604.02834) · MedHall-Bench (medical hallucination) · MedHarm-Bench (harmful medical advice). Reproduce any of them with one command via mirobody-eval, which also seeds a deployment with synthetic, PHI-free trajectories.

Run the whole thing

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

--depth 1 because the history is mostly superseded frontend builds and you almost certainly do not want it: it takes the clone from ~125 MB to ~99 MB (measured 2026-09-14 — most of what is left is the LFS bundle, which both copies need). Drop the flag if you plan to send a pull request. ./deploy.sh also fetches the 22 MB concept graph behind semantic indicator search, which is a release asset rather than a tracked file — see mirobody/res/EXTERNAL.tsv.

Sign in as [email protected], code 111111. No mail provider needed: the sign-in page opens on password, and an account of your own is one request away:

curl -X POST localhost:18060/password/register -H 'Content-Type: application/json' \
     -d '{"email":"[email protected]","password":"at-least-8-chars"}'

Your care circle. The account is named for the role it plays: you sign in as the caregiver, and the record you read belongs to someone else. A circle is the unit of sharing — each member holds their own record and decides, on their own row, whether the others may view it.

SEED_DEMO_DATA is on by default, so the circle is already populated: you own a thin record — a few weeks of self-tracked vitals and one unremarkable checkup — and one synthetic person shares a thick one with you, Demo (synthetic): 244 indicators, 14,273 readings across two years, five documents the agent can read. Same question, two records: your HbA1c answers with one boring-normal value from data you own; hers with a two-year story from data you can only view. The recording walks both sides — your own indicators and files, then the switch to her shared record and its two years of HbA1c:

The switch in that diagram is a column, not a promise: care_circle_members.health_access, NOT NULL DEFAULT 0, on your own row. Being invited into a circle shares nothing — the member decides, and no other person's action can raise it. A route that forgets to check answers 403 instead of handing over a record. examples/06_care_circle_rules.py prints the whole decision table offline. Set SEED_DEMO_DATA=false for a deployment that will hold real data.

One key runs everything, and none of it is your GPU. The key points at a hosted model; your machine runs no inference. Browsing the seeded record needs no key; the upload and the questions below ride one. Put ONE key in the .env next to compose.yaml and docker compose restart — the app re-reads /app/.env; a shell export does not reach the containers. The .env is the only place for the key: config.llm.yaml names the variable (api_key: OPENROUTER_API_KEY), never the secret. An OpenRouter key as OPENROUTER_API_KEY is the recommended one; a DashScope key when openrouter.ai is unreachable from your network; a Google, OpenAI (OPENAI_API_KEY), Anthropic (ANTHROPIC_API_KEY) or DeepSeek key works alone as well. Every model decision — which model chats, which reads report photos, which extracts the indicators, which embeds — is a line in config.llm.yaml, where you can read and change it (a self-hosted gateway is one _BASE_URL in .env). The boot log, and mirobody doctor, print what each surface selected and name the fix where one has nothing.

① Collect + ② Translate. Drop demo/lab_report_2025-10-15.pdf on the Data page and twelve analytes come out with values and units, each linking back to the page it was read from:

③ Agent. Ask about her HbA1c and the agent finds the data, charts the three lab draws against 104 sensor-derived estimates, and says plainly that the improvement did not hold. Ask again about the report you just uploaded and it reads that instead — the fourth scene of the four-minute walkthrough.

→ Docker deployment · Configuration · Local Python setup

Use it, extend it

Contributing

The highest-leverage contribution is a term the resolver gets wrong. Run mirobody resolve ""; 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 · Contributing guide · Repository layout · Roadmap · CHANGELOG · SECURITY

Documentation

docs.mirobody.ai, in English and Chinese — start at the [Quickst