How miromindai/miroflow works
Project type and purpose
• MiroFlow is a modular AI‐agent framework designed to orchestrate complex, multi-tool tasks using large language models (LLMs).
• It provides:
• A core pipeline & orchestrator for multi‐turn conversations and error recovery
• A unified LLM client layer supporting multiple providers
• A tool‐manager that connects to lightweight microservices (“MCP servers”) for external capabilities (code execution, vision, search, audio, etc.)
• A suite of CLI applications for benchmarking, evaluation, and trace visualization
• Config management (Hydra/YAML), logging/tracing, and a simple web UI for visual debugging.
File structure
Top‐level directories:
• apps/ — four CLI/web apps:
– eval-agent: run evaluation pipelines programmatically
– prepare-benchmark: download/generate benchmark data
– run-agent: main entrypoint for executing tasks & benchmarks
• libs/miroflow-contrib — optional tracing & contrib modules
• libs/miroflow-tool — tool manager + MCP server implementations (FastMCP microservices)
• libs/miroflow — core framework: pipeline, orchestrator, LLM client, config, logging, utils
• data/ — raw or generated datasets
• docs/ — figures, Dockerfile, contribution docs
Patterns & deployment files:
• justfile, .github/workflows — CI/CD, lint/test automation
• REUSE.toml, LICENSES — licensing compliance
• Each app has a pyproject.toml & uv.lock for reproducible environments
Key README insights
• Describes five major stages: intent recognition, planning/orchestration, sub-agent delegation, tool access via MCP, result synthesis
• Enumerates core modules under libs/miroflow and libs/miroflow-tool