Mirobody

[Asking how cholesterol has changed: the agent shows each step as it works, charts three reports that name the test differently as one trend, and clicking an underlined number opens the reading and the file it came from]

Three reports, three names for the same test. One trend, with the file behind every number.

Choose where to start

Try the engine with one command

With uv installed, paste this into your terminal:

uvx --python 3.12 mirobody resolve "LDL cholesterol" 血红蛋白 ヘモグロビン "Total Kolesterol" 血脂

No repository clone, Docker, model or API key. uv prepares Python and the package for you; the first run needs a network connection to download them. The resolver then works offline from the bundled vocabulary.

血红蛋白 and ヘモグロビン both resolve to 718-7. 血脂 stays unresolved: it names a category, not one test. The same engine standardizes names and units across the full app.

[mirobody resolve: Chinese and Japanese names for haemoglobin resolve to the same LOINC code, while a broad category stays unresolved]

Give the same engine to your AI agent. In Claude Code:

claude mcp add mirobody -- uvx --python 3.12 mirobody mcp

Your agent can then standardize any name, reading, unit or complaint you paste, on your machine and without a key. Reading a whole report also needs mirobody[parse] and a model key. Other MCP clients take the same command.

→ Use the engine from Python or MCP

Try the full app in Codespaces

Run the web app in your browser with your own model API key. GitHub prepares the environment; you do not need to install Docker or Python on your computer.

[Open in GitHub Codespaces]

1. Create the codespace and wait for the automatic setup to finish in its terminal. It starts the app, Postgres, the worker and the synthetic demo data.
2. Open the setup link printed in the terminal and enter your own model API key: OpenRouter, OpenAI, Gemini, Anthropic, DeepSeek, DashScope, or any OpenAI-compatible gateway. The key is saved only after a real request succeeds. The forwarded app port is private by default.
3. Sign in as [email protected], code 111111, on the Email code tab. Ask "How has my cholesterol changed?" The demo holds 2,019 readings across your record and [email protected]'s, shared with you view-only.
4. Upload a synthetic sample from demo/upload/ on the Data page to see a new report join the record, or switch to mom's record and ask again.

The app and its records run in GitHub's cloud; the model service receives the documents and questions it processes. Use the included synthetic data to explore. Codespaces has usage limits and billing; model usage is billed by your model provider. Stop the codespace when you finish. For complete control on your own machine, use the local route below.

Run everything 100% locally

The app, records and models all run on your own machine. No model API key needed. Docker with Compose starts the stack and llama.cpp, with a small answering model and a document reader:

git clone --depth 1 https://github.com/thetahealth/mirobody.git && cd mirobodyCOMPOSE_PROFILES=local-cpu ./deploy.sh

Allow 16 GB of memory, at least 8 GB of it for Docker. No GPU is required; on Windows, run the commands in a WSL 2 terminal. The first start needs a network connection to download container images and model weights. If the registry is not reachable from your network, deploy.sh says so and retries through a mirror; DOCKER_MIRROR in .env names your own. After those downloads, document reading, standardization and answers run locally. CPU answers can take minutes; Mac, NVIDIA and Windows options explain how to use your GPU.

Open the app link the script prints. The local models are already selected; sign in with the same demo account above, ask a question or upload a sample. Which model to choose compares the local options.

[The first-run page: a model name edited beside an OpenRouter key, then 100% on this machine: the page finds the llama.cpp server, lists the models it serves, and both models are ready]

Prefer Docker with a cloud model? Run ./deploy.sh without a local model profile, then paste your own model key into the printed setup link. The app and records stay on your host; the model service processes your documents and questions.

Chat calculations take the original queried data directly. A cloud model's code runs as JavaScript inside the app, with nothing to set up; a model on your own hardware computes in Python in a local Docker container when the app can reach a Docker daemon. Setup and data flow cover both. Query views provide their built-in statistics without either.

→ Self-host guide · Walkthrough · Deploy on a server

Run without Docker

No container, and nothing pulled from a registry. On Debian or Ubuntu, PostgreSQL 16 with pgvector comes from the distribution's own packages and the app runs straight from the checkout. It is the same install the Codespaces path performs:

git clone --depth 1 https://github.com/thetahealth/mirobody.git && cd mirobodyscripts/install_native.sh   # PostgreSQL, pgvector, the Python package, a .envscripts/start_native.sh     # the worker and the app on port 18060

Open the link it prints and sign in as [email protected], code 111111: the demo record is seeded already. A model key from the printed setup link adds document extraction and answers. install_native.sh installs system packages with sudo and writes one sudoers rule for the postgres account, and says so as it runs.

The local models belong to the Docker path, so this one answers through a model service. Everything else matches the row above it in what stays on your machine: the records live in the PostgreSQL you just installed, and a name to a code never leaves the machine.

What you can do

• Compare reports across labs. PDFs, phone photos and spreadsheets, 23 file types in all, land in one history. Every reading links to the page it was read from.
• Bring in everyday data. Import an Apple Health export, or connect Garmin, Oura and Whoop.
• Look after your family. Invite someone who shares their own record, or keep one for a parent who never signs in.
• Write a health journal. headache since last night, BP 150/95, no fever becomes a coded complaint and two coded readings; "no fever" is not logged as a fever.
• Explore genetic data. Upload a 23andMe, AncestryDNA or WeGene export, or a VCF, and ask by rsID, gene or region. Drug questions get the matching CPIC guideline coverage, never a change of medication.
• Use your own agent. Connect Claude Code, Codex or Cursor over MCP, or embed the offline engine in Python.

Mirobody does not diagnose or prescribe, and its answers are not medical advice.

Collect · Standardize · Agent

[Collect reports, device data and genetics; standardize names and units so sources compare; ask across your history and trace each answer to its source]

The model reads documents and reasons; the codes and units come from vocabularies bundled with the package, never from the model. How a reading is standardized, step by step · How a reading moves through the pipeline.

Care for your family

[Invite a family member; they choose whether you may view or edit their own record; ask about the record they share. A family member who does not sign in can be managed by you until they take over and choose your access]

Joining a care circle shares nothing by itself. Each member decides whether others may view or edit their own record. A parent who never signs in can be managed by you until they claim the record and choose what you keep. See it in the walkthrough.

What stays on your machine

Where the app runs and where the model runs are separate choices. Mirobody does not report usage anywhere.

The command-line resolver also runs on your machine without a model or network access after its initial download.

Device vendors see data only after you link one. Before this reaches a network you do not control, read SECURITY.md: it lists every address the server can call.

Build on Mirobody

Over MCP the stack serves tools scoped to the signed-in person: four read your record (readings, medications, genotypes, pharmacogenomics), three resolve names and units, and two search medical reference material when the deployment has a knowledge source.

The stdio server is published to the official MCP Registry as ai.thetahealth/mirobody with every release. Directories that read the registry list it too, such as mcprush and AddonAudit, which runs it in a sandbox. Each picks up a new version on its own schedule, so take the install line from this README.

After pip install mirobody, use the same engine from Python:

from mirobody.engine import resolve, resolve_reading
resolve("血红蛋白").loinc                                     # '718-7'   any language, one coderesolve_reading("total cholesterol", "5.0", "mmol/L").loinc  # '14647-2' the unit picks the code...resolve_reading("total cholesterol", "193", "mg/dL").loinc   # '2093-3'  ...mass, not molesresolve("中性粒细胞百分比").loinc                               # '26511-6' Neutrophils/Leukocytesresolve("血脂").resolved                                     # False    a category, not one test

Numbers you can check

Mirobody is the health data engine behind Theta, a live consumer health app on the web and mobile, with these users today:
[Theta registered users, live] [Theta daily active users, live]

Contributing

The most useful contribution is a term the resolver gets wrong. Run mirobody resolve "<term>"; if the answer is wrong or empty, report it or send a fix with a test case. That is how the vocabulary grows into new languages: Russian panel and Estonian vitamin terms arrived in #88 and Turkish lab-report spellings in #140, each one row of resolver_overrides.tsv and one test case per term. CONTRIBUTING.md has the setup and the checks.

Built on HL7 FHIR, LOINC from the Regenstrief Institute, UCUM, ICPC-3 (WONCA), CPIC, llama.cpp and deepagents, with thanks. Terminology licences: LICENSE-3RD-PARTY.