🚀 Mirobody

Your Data, Your AI — Health, Finance & More. Open Source, Privacy-First.

Self-hosted data platform that bridges your personal data with the latest AI capabilities

AI Engine:

• 🌐 HTTP Remote MCP Server - Deploy and access MCP tools over HTTPS
• 🎯 Claude Agent Skills Support - Develop tools using standard Skills format (SKILL.md)
• 🔄 Universal Tool Adapter - Works with ChatGPT, Claude, Cursor, and more
• 🔌 Pluggable Data Providers - Connect any data source via Providers API
• 🤖 Custom Agents - Create your own conversational agents via Agents API

Health Data:

• 🏥 FHIR & Health Standards - 400+ health indicators, LOINC / SNOMED CT / RxNorm cross-vocabulary search
• 📊 Health Data Pipeline - Ingest from 300+ wearables, Apple Health, EHR; normalize to StandardPulseData

📖 Table of Contents

• Why Mirobody?
• Architecture — AI & Agent Engine, FHIR & Health Standards, Health Data Pipeline, Infrastructure
• Theta Wellness: Our Health Intelligence App
• Quick Start
• Access & Authentication
• API Reference
• Documentation

✨ Why Mirobody?

🔄 Write Tools Once, Run Everywhere

Forget about complex JSON schemas, manual bindings, or router configurations. In Mirobody, your Python code is the only definition required.

• Tools built here instantly work in ChatGPT (via Apps-SDK) and the entire MCP Ecosystem (Claude, Cursor, IDEs).
• Mirobody works simultaneously as an MCP Client (to use tools) and an OAuth-enabled MCP Server (to provide data), creating a complete data loop.
• 🌐 HTTP Remote MCP Support: Mirobody supports HTTP-based remote MCP servers, enabling cloud deployments and cross-network tool access. Configure MCP_PUBLIC_URL to expose your MCP server over HTTPS for ChatGPT Apps and other remote integrations.

💎 Your Data Is an Asset, Not a Payload

Mirobody is built for Personal Intelligence, not just local storage. We believe the next frontier of AI is not knowing more about the world, but knowing more about you.

• General AI creates generic answers. Mirobody uses your data to create a Personal Knowledge Base, enabling AI to give answers that are truly relevant to your life.
• You can run the entire engine locally on your machine. We provide the architecture to unlock your data's value without ever compromising ownership.

🤖 Native Agent Engine

• Powered by a self-developed agent engine that fully reproduces Claude Code's autonomous capabilities locally.
• 🎯 Skills-Based Tool Development: Mirobody supports developing tools using Claude Agent Skills format (SKILL.md files). You can create reusable tools that work seamlessly across the MCP ecosystem. Simply structure your tools as Skills and drop them into the skills/ directory - Mirobody will automatically discover and expose them.
• Designed to load Claude Agent Skills SKILL.md files, turning your private data into an actionable knowledge base.

🧠 Agent Architecture

Mirobody provides three agent types for different use cases:

| Agent | Description | Use Case | | ----------------------- | ------------------------------- | ------------------------------------------------------ | | DeepAgent | Single-model tool orchestration | Complex queries requiring data retrieval and analysis | | MixAgent | Two-phase model fusion | Optimized cost/quality balance with specialized models | | BaselineAgent | Direct LLM conversation | Simple Q&A without tool calls |

DeepAgent

Inspired by LangChain DeepAgents, DeepAgent is our primary agent for tool-assisted conversations. Key features:

• Full MCP Tool Support: All tools are MCP-compliant and configurable via ALLOWED_TOOLS / DISALLOWED_TOOLS
• Multi-Provider: Supports Google GenAI, OpenAI, Anthropic, and OpenRouter
• Middleware Stack: Includes prompt caching, tool call patching, and message summarization

MixAgent

A two-phase model fusion architecture that separates tool orchestration from response generation:

• Phase 1 (Orchestrator): A capable model (e.g., Claude Sonnet) handles tool calls and data collection
• Phase 2 (Responder): A cost-effective model (e.g., Gemini Flash) generates the final response with collected context

This architecture optimizes for both cost and quality by using expensive models only where necessary.

BaselineAgent

A lightweight agent for direct LLM conversations without tool access. Ideal for:

• Simple Q&A scenarios
• Testing and development
• Low-latency responses

📁 Secure File Operations: File tools (ls, read_file, write_file, edit_file, glob, grep) are backed by PostgreSQL for data persistence and auditability. See mirobody/pub/tools/file_read_service.py and file_write_service.py for implementation details.

🧪 Sandbox Code Execution: The execute tool runs shell commands in isolated E2B cloud sandboxes for data analysis and computation. Requires E2B_API_KEY configuration. See CONFIG for setup details.

👉 See CONFIG for detailed agent configuration guide.

🏗️ Architecture

AI & Agent Engine

| Module | Path | Description | |--------|------|-------------| | Chat Service | mirobody/chat/ | Session management, conversation history, streaming adapters (HTTP/WebSocket), memory integration | | Agent Implementations | mirobody/pub/agents/ | DeepAgent (LangChain), MixAgent (two-phase fusion), BaselineAgent | | LLM Clients | mirobody/utils/llm/ | Multi-provider adapter (OpenAI, Gemini, Azure OpenAI, Volcengine, Dashscope), HIPAA-compliant routing | | MCP Server | mirobody/mcp/ | JSON-RPC 2.0 tool/resource server, local + HTTP remote access | | Tools | mirobody/pub/tools/ | Built-in tools: file ops, charts, code execution (E2B sandbox), memory | | Embeddings | mirobody/utils/embedding.py | Provider-agnostic 1024-dim embeddings (Gemini / Qwen), pgvector semantic search | | Prompt Templates | prompts/ | Jinja2 system prompts with dynamic context injection (user timezone, tools, health profile) | | Skills | skills/ | Claude Agent Skills (SKILL.md + metadata.json), auto-discovery |

FHIR & Health Standards

| Module | Path | Description | |--------|------|-------------| | FHIR Mapping | mirobody/pulse/core/fhir_mapping.py | In-memory cache of indicator → FHIR code, optional auto-registration of new codes | | Indicator Registry | mirobody/pulse/core/indicators_info.py | 400+ StandardIndicator enum, multi-source (Vital, Apple Health, Garmin, Whoop, Renpho) | | Unit Conversion | mirobody/pulse/core/units.py | Bidirectional conversion: kg/lbs, °C/°F, mg·dL⁻¹/mmol·L⁻¹, mmHg/kPa, etc. | | Indicator Search | mirobody/indicator/ | Embedding-based free-text → indicator code, concept graph expansion (LOINC / SNOMED CT / RxNorm bridges) | | Medical Code Mapping | health_tools/ | SNOMED-CT code mapping, health indicator classification |

Health Data Pipeline (Pulse)

| Module | Path | Description | |--------|------|-------------| | Platform Manager | mirobody/pulse/ | Platform–Provider plugin architecture, data normalization to StandardPulseData | | Theta Platform | mirobody/pulse/theta/ | Direct device integrations: Garmin, Whoop, Oura, Renpho, PostgreSQL | | Apple Health | mirobody/pulse/apple/ | Apple Health import, CDA (Clinical Document Architecture) processing | | Data Upload | mirobody/pulse/data_upload/ | StandardPulseData → th_series_data write pipeline | | File Parser | mirobody/pulse/file_parser/ | Multi-format: PDF, CSV, Excel, audio, image, genetic data; LLM-powered indicator extraction | | Aggregation | mirobody/pulse/core/aggregate_indicator/ | Series → daily summaries, derived metrics, sleep 18:00–18:00 window | | Health Insights | mirobody/pulse/core/insight/ | AI-powered trend detection, anomaly analysis, pattern recipes (multi-signal, recovery, glucose) |

Infrastructure

| Module | Path | Description | |--------|------|-------------| | Configuration | mirobody/utils/config/ | YAML + env var layered config, Fernet encryption, multi-storage backend (Local / S3 / Aliyun OSS) | | Auth & User | mirobody/user/ | JWT, OAuth (Google / Apple), WebAuthn / FIDO2, email verification | | Server | mirobody/server/ | Starlette ASGI, JWT middleware, rate limiting | | Database | mirobody/utils/db.py | Async PostgreSQL (psycopg), Redis cache/session store |

Extension Points (Root Directories)

| Directory | Purpose | |-----------|---------| | tools/ | Drop-in Python tools — auto-discovered as MCP tools | | skills/ | Claude Agent Skills (SKILL.md + metadata.json) | | agents/ | Custom agent implementations | | providers/ | Custom Theta data providers | | prompts/ | Jinja2 prompt templates | | resources/ | Static resour