Hanzo Flow
Build an AI workflow on a canvas, ship it as an API
Drag a model, a vector store and a prompt onto a canvas, wire them together, and run it with the values you actually have. Every component is Python you can open and edit in place, so the canvas stops where your code starts instead of at a wall. When it works, publish it: a REST endpoint, a JSON file a Python app loads, or an MCP server, which turns the flow into a tool any MCP client can call.
The canvas is not the ceiling
Visual builders usually work right up until you need the thing they did not think of. This one hands you the source.
The library
Models, vector stores, document loaders, text splitters, output parsers, memory, retrievers and toolkits — plus anything you write yourself, which is a component like the rest.
Open any node
A component is a Python class. Edit it in the browser and the change is live on the next run — no rebuild, no plugin format, no waiting for someone to add the parameter you need.
Step through it
The playground runs a flow one node at a time and shows what each one received and returned. When a chain gives a strange answer, this is where you find the node that caused it.
Multi-agent
Several agents on the same canvas, with conversation state between them and retrieval underneath — visible as a graph rather than buried in a prompt.
Publish it three ways
As a REST endpoint, as JSON your Python app loads directly, or as an MCP server — after which the flow is a tool your coding agent can call.
Keys stay out of the flow
Credentials live as variables the graph references by name, so a flow you export or share carries the wiring and not the secrets.
Up to 5% of compute goes back to open source
Every deployment is SBOM-verified. Contributors to Langflow earn a share of compute revenue — transparent, on-chain, and customizable by the community.
Run it locally first
uv pip install hanzoflow, then uv run hanzoflow run. It opens on 127.0.0.1:7860. Docker works too.