Langflow

Langflow lets you build an LLM app, such as an agent with tools, a RAG chatbot or a multi-step pipeline, by connecting components on a canvas instead of writing the orchestration code. It suits prototyping quickly and letting people who don't write the app's code change prompts and flows, while your product calls the result over HTTP.

How it works: a flow is a graph of components, from inputs, outputs and data stores to agents, language models and embedding providers. When the built-in set isn't enough, you write a custom component as a Python class, and its inputs and outputs show up as a node in the visual editor. A playground lets you test a flow as you build it.

What you get: every flow can be run from your code through the /api/v1/run endpoint with a Langflow API key, with "tweaks" to override component settings per request. Flows can also be triggered by webhooks, embedded in a site as a chat widget, called through an OpenAI-compatible responses endpoint, or exposed as tools on Langflow's MCP server. You can run it as Langflow Desktop on macOS or Windows, as a Python package, or in Docker, and flows export as JSON files you can import into another instance.

The limits: the flow logic lives in Langflow's database as JSON rather than as code in your repository. The default database is SQLite, and the docs recommend PostgreSQL for production. Langflow is built to execute developer-provided code with full access to its host, and it does not isolate users from one another, so a shared or public deployment needs isolation at the infrastructure level.