ToolFlow

中文 | English

---

Stop building workflows. Write a function, ship a tool.

ToolFlow is a runtime-first MCP tool system for LLM applications. You write plain Python functions — ToolFlow handles versioning, dispatch, isolation, and dynamic composition at runtime.

No DAGs. No pipelines. No restarts. Tools are loaded, executed, and destroyed on demand in isolated sandboxes. Compose them dynamically when called, not when deployed.

Built on Django (control plane) + FastMCP (stateless execution layer).

• 🐍 Python-native tools — write a function, register a tool. No boilerplate.
• ⚡ Dynamic composition — tools combine at runtime, not at design time.
• 🔒 Ephemeral sandboxes — every call runs in a fresh in-process sandbox with restricted imports; tool state is not reused across calls, though external side effects (for example, network calls) are still possible.
• 🧱 Control plane / execution plane separation — Django manages assets; FastMCP executes statelessly.
• 🔄 Built-in lifecycle management — version, release, and monitor tools independently.

graph TD
  A[LLM / Client] -->|Tool Call| B[Django Gateway + Control Plane]
  B -->|Fetch Tool Code/Version| C["DB / Storage"]
  B -->|Dispatch Tasks| D{Queue/Dispatcher}
  D --> E[FastMCP Executor 1]
  D --> F[FastMCP Executor N]
  E -->|Ephemeral Load + Execute| G[Ephemeral Tool Runtime Sandbox]
  F -->|Ephemeral Load + Execute| H[Ephemeral Tool Runtime Sandbox]
  E -->|Report Status/Result| B
  F -->|Report Status/Result| B

• server/: Django gateway and management APIs
• runtime/: executor, bridge service, and runtime config
• frontend/: React + Vite frontend
• start_services.py: one-command local orchestrator

1. Set up Python environment

python -m venv .venv
.