Core Concepts
MiroFlow is a flexible framework for building and deploying intelligent agents capable of complex reasoning and tool use.
Architecture Overview
Multi-Stage Agentic Process
MiroFlow processes user queries through a structured workflow:
- Intent Recognition & Query Augmentation - LLMs analyze and refine user input
- Planning & Task Orchestration - Main agent creates execution plans and coordinates sub-agents
- Delegation to Sub-Agents - Specialized agents handle domain-specific tasks
- Tool Access via MCP Servers - Agents leverage external capabilities through MCP protocol
- Result Synthesis & Output Alignment - Final results are synthesized and formatted
Core Components
Agent System
Agent Architecture
Main Agent: Primary coordinator that receives tasks, creates plans, and manages overall execution. Can use reasoning tools and delegate to sub-agents.
Sub-Agents: Specialized agents for specific domains:
agent-worker: General-purpose agent with comprehensive tool access (search, files, code, media)- Each sub-agent has dedicated configurations and can operate independently
Tool Integration
Tool System
Tool Manager: Connects to MCP servers and manages tool availability
Available Tools:
- Code Execution: Python sandbox via E2B integration
- Web Search: Google search with content retrieval
- Document Processing: Multi-format file reading and analysis
- Visual Processing: Image and video analysis
- Audio Processing: Transcription and audio analysis
- Enhanced Reasoning: Advanced reasoning via high-quality LLMs
See Tool Overview for detailed tool configurations and capabilities.
LLM Support
Multi-Provider Support
Unified interface supporting:
- Anthropic Claude (via Anthropic API, OpenRouter)
- OpenAI GPT (via OpenAI API)
- Qwen (via SGLang)
- MiroThinker (via SGLang)
- see LLM Clients Overview for details*
Documentation Info
Last Updated: September 2025 · Doc Contributor: Team @ MiroMind AI