The full-stack kit for AI agents that actually know your body

What it does

Mirobody is a self-hosted Python stack that ingests data from over 300 wearables, Apple Health, EHRs, and genetic files, then normalizes everything into a standard schema called StandardPulseData. On top of that data layer it runs a native agent engine—complete with an MCP server, multi-provider LLM routing, and a skills/ directory—so you can ask questions about your own biomarkers instead of getting generic advice. The project ships with FHIR mappings, LOINC/SNOMED CT bridges, and unit conversions baked in, which is the unglamorous but essential plumbing that keeps the AI from confusing milligrams with millimoles.

The interesting bit

The architecture treats your health data as an MCP resource: the engine acts as both an MCP client and an OAuth-enabled MCP server, meaning it can consume external tools while exposing your personal data to ChatGPT, Claude, or Cursor without a middleman. Its MixAgent even splits work between an expensive orchestrator model and a cheap responder model to keep API costs sane during long sessions.

Key highlights

• MCP-native by design: Tools are auto-discovered from Python files and SKILL.md entries, then exposed via JSON-RPC 2.0 locally or over HTTPS.
• Serious health plumbing: 400+ standard indicators, embedding-based free-text search across medical vocabularies, and LLM-powered extraction from PDFs, images, and genetic data formats.
• Cost-cognitive agents: MixAgent uses a two-phase model fusion—capable model for tool calls, cheaper model for response generation—to balance accuracy and spend.
• Privacy-first infrastructure: Runs fully offline with JWT/OAuth/WebAuthn auth, PostgreSQL-backed file operations, and optional E2B sandboxes for code execution.
• Generalizable beyond health: The authors bill it as a vertical-agnostic engine; swap the tools/ directory and the same agent infrastructure purportedly works for finance or legal stacks.

Caveats

• The README pitches easy vertical expansion—just swap the tools/ directory—but the architecture tables reveal deep entanglement with FHIR, LOINC, and lab-specific unit conversions that look hard to generalize.
• Several features, such as E2B sandbox execution and remote MCP HTTPS access, require external API keys or cloud services, so “fully offline” depends on which knobs you turn on.
• The README repeats the claim that the flagship app covers EHR systems for “90% of the US population” without offering supporting detail.

Verdict

If you are building a personal-health AI product and need a self-hosted backbone that speaks FHIR and MCP, this is a rare head start. If you are looking for a lightweight drop-in health SDK, the sheer surface area—agents, auth, vector search, and ASGI server included—might feel like adopting an entire operating system.