FLUJO
by punkpeye
A local-first visual builder for chaining models and MCP tools into automated workflows, with a built-in debugger and an MCP proxy for re-exposing flows to other clients.
About
FLUJO is a desktop app for wiring together language models, MCP servers, and reusable subflows on a drag-and-drop canvas rather than in code. Each node in a flow can call a different model (OpenAI, Anthropic, Gemini, or a local Ollama model) and can be given access to specific installed MCP tools, so a single workflow might route a task through a fast model for triage and a stronger one for the actual work, calling tools in between.
Beyond building flows, FLUJO manages API keys and environment variables centrally so credentials aren't scattered across config files, can install MCP servers straight from a GitHub URL, and includes a step-through debugger for watching exactly what a flow sends and receives at each node. It can also act as an MCP proxy, re-exposing any server it has installed to external clients like Claude Desktop or Cursor over Streamable HTTP, and it serves finished flows through OpenAI-compatible and MCP endpoints so other apps can trigger them directly.
Key features
- Drag-and-drop flow builder connecting models, MCP tools, and reusable subflows
- Installs MCP servers directly from a GitHub repository URL
- Centralized, encrypted-at-rest storage for API keys and environment variables
- Step-through debugger for inspecting each node's input/output during a run
- Acts as an MCP proxy, re-exposing installed servers to other MCP clients
- Automation triggers (scheduled runs, webhooks, file watching) plus a chat interface for testing flows
Use cases
- Building a multi-model pipeline (cheap model for routing, capable model for execution) without writing orchestration code
- Centralizing and reusing MCP server installs across several local AI clients via the proxy feature
- Debugging why a tool-calling workflow produced a bad result, step by step
- Exposing a finished internal workflow to other tools as an OpenAI-compatible or MCP endpoint
Available tools
Flow builder
Visual canvas for connecting model nodes, MCP tool nodes, and subflows into a runnable pipeline.
MCP proxy
Re-exposes any MCP server installed in FLUJO to external clients like Claude Desktop or Cursor.
Debugger
Steps through a flow's execution node by node, showing inputs and outputs at each stage.
Frequently asked questions
Can FLUJO run models offline?
Yes. Nodes can call a local Ollama model instead of a hosted API, so a flow can run without sending data to an external provider.
Can other apps trigger a FLUJO workflow directly?
Yes. Finished flows are served through OpenAI-compatible and MCP endpoints, so tools like Cline, Roo, or Claude can call them like any other model or tool.