# mirobody [Health: Active]

**Category:** 🧬 Biology & Bioinformatics  
**Repository:** https://github.com/thetahealth/mirobody  
**GitHub Stars:** 1381  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/mirobody

## Description
Turn scattered health data into answers you can trace. Self-hosted AI health data engine for lab reports, wearables and genetics, with a built-in MCP server. Run it with Docker and one LLM key, or keep every model local on llama.cpp.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "mirobody": {
    "command": "uvx",
    "args": ["--python","3.12","mirobody"]
  }
}
```

## Documentation

## What mirobody does

The mirobody MCP server exposes Mirobody's health-data engine to compatible agents. Mirobody combines laboratory reports, wearable imports, genetic files, and journal entries into a standardized record. The system can compare readings across reports, preserve links to the source page for each value, and answer questions about trends such as changes in cholesterol.

Health records can include PDFs, phone photos, spreadsheets, Apple Health exports, and connections to Garmin, Oura, and Whoop. Genetic data can come from 23andMe, AncestryDNA, WeGene, or VCF files. A journal entry can also be converted into coded complaints and readings while distinguishing negated statements, such as “no fever.”

This places the project in the Biology & Bioinformatics category, although its inputs also cover general personal health and activity data. Genetic questions can be asked by rsID, gene, or region. Drug-related genetic questions use CPIC coverage and do not produce medication-change instructions.

## How it works

Mirobody first translates incoming health information into a common record. Its resolver maps different names and languages for the same test to standardized identifiers and units; for example, multiple names for haemoglobin can resolve to LOINC code 718-7. Broad categories that do not identify one test remain unresolved.

The application stores records in PostgreSQL with pgvector in the documented self-hosted deployments. A worker handles background processing, while the application presents the resulting record and answers. Each extracted reading can retain the document page it came from, allowing an agent or user to trace a result back to its source.

The mirobody MCP server is intended for agent access through Claude Code, Codex, or Cursor. The underlying engine can also be used from Python. Model execution is separate from the record-processing components: a cloud model service can process documents and questions, or llama.cpp can run the models locally.

## Setup and configuration

The repository documents three deployment paths. The Docker Compose route starts the app, database, worker, and either local llama.cpp models or an application configured with a model provider. The local CPU route needs Docker Compose, about 16 GB of memory, and at least 8 GB allocated to Docker; no GPU or model API key is required for the fully local profile.

A native installation is available for Debian or Ubuntu without Docker. It uses PostgreSQL 16 with pgvector, Python 3.12, and the repository’s installation and startup scripts. The Codespaces route runs the application in GitHub’s cloud and requires a model API key for the full application.

Supported model-key options listed in the README include OpenRouter, OpenAI, Gemini, Anthropic, DeepSeek, DashScope, and OpenAI-compatible gateways. The provided material does not specify a separate MCP installation command or required MCP environment variables.

## Tools and capabilities

- Resolve health-test names and units through the bundled vocabulary, including multilingual names.
- Compare laboratory results across reports and preserve source-file references.
- Import Apple Health data and connect Garmin, Oura, and Whoop records.
- Parse journal text into coded complaints and readings.
- Query genetic files by rsID, gene, or genomic region.
- Ask agent questions over MCP from Claude Code, Codex, or Cursor.
- Use the engine from Python without requiring a model for vocabulary resolution.

## Limitations and notes

The one-command resolver works offline after its package and Python dependencies have been downloaded, but the first run needs network access. Local model deployments also need an initial connection to download container images and model weights. CPU-based answers may take minutes.

Cloud-model deployments keep the app and records on the selected host, but the model service processes the documents and questions sent to it. The fully local Docker profile keeps the app, records, and models on the user’s machine. The Codespaces demonstration uses synthetic data and has GitHub Codespaces and model-provider usage costs.

Genetic drug questions provide CPIC coverage but are explicitly not a basis for changing medication. The supplied documentation names supported MCP clients but does not provide detailed MCP tool schemas or client-specific configuration examples.

_Full upstream README: https://allmcps.com/mcp/mirobody/readme_