Flowise

Where Flowise sits

Flowise is the visual builder I recommend when someone wants LangChain-grade control without writing LangChain code. The open-source project (FlowiseAI/Flowise on GitHub) was created in March 2023 and grew fast — roughly 55,000 stars and nearly 25,000 forks as I read the repository in October 2026 — which makes it one of the most-used LLM orchestration projects out there. It is built on LangChain.js underneath, ships as an npm package or Docker image, and offers a hosted cloud alongside the self-host path. The pitch on the homepage is deliberately modest: modular building blocks for anything from simple compositional workflows to autonomous agents, iterating fast.

The licensing deserves precision, because "open source" here has an asterisk. Flowise's LICENSE.md splits the codebase: content outside the enterprise directory is Apache 2.0, while the server's enterprise directory and explicitly noted files (identity management and similar) carry a commercial license. In practice, the core builder, flows, chatflows, agentflows, and deployment tooling are free to self-host; the enterprise layer — the identity, permissioning, and admin features organisations eventually ask for — is the paid part. That is a fair split and it is stated openly in the repo, but buyers who assume the whole repo is Apache 2.0 will be surprised at procurement time.

What it actually does

The product is a node canvas. You drag components — chat models, embeddings, vector databases, document loaders, tools, memory, agents, chains — connect their ports, and get a runnable LLM application. The current feature set (from flowiseai.com and the docs) has grown well past simple chatbots: Agentflow for multi-agent orchestration, human-in-the-loop steps where a person reviews an agent's work mid-flow, evaluations and metrics, observability into runs, and deployment targets ranging from embedded chatbots to REST APIs. The vendor claims support for more than 100 LLMs, embedding models, and vector databases, with on-prem and cloud deployment and horizontal scaling via a message queue and workers. For grounding work, you wire a document loader to a splitter to an embedding model to a vector store to a retrieval node — each step visible and replaceable — which is exactly the transparency that prompt-only platforms hide.

Pricing for the hosted cloud at time of writing, from flowiseai.com in October 2026: Free at $0 with 2 flows and assistants, 100 predictions per month, 5 MB storage, evaluations and metrics, custom embedded chatbot branding, and community support. Starter at $35 per month with unlimited flows and assistants, 10,000 predictions, 1 GB storage, and a first-month-free offer. Pro at $65 per month with 50,000 predictions, 10 GB storage, unlimited workspaces, 5 users plus $15 per additional user per month, admin roles and permissions, and priority support. Enterprise is custom (the page positions it for on-prem and enterprise-grade infrastructure). Self-hosting is free under the license terms above; what you pay for there is your own infrastructure and the enterprise features if you want them.

My testing notes

Four observations from building with it.

First, the canvas is genuinely fast for prototyping. Standing up a retrieval agent — loader, splitter, embeddings, vector store, conversational agent with a tool — took me one sitting, and the instant-predict box on each flow means you see what a node produces before you commit to the downstream wiring. When something misbehaves, the answer is usually visible: a splitter config producing garbage chunks, a retriever returning four irrelevant documents. In my testing the visual layer did not hide errors so much as point at them, which is what I want from a builder.

Second, the gap between "runs on my laptop" and "runs in production" is where the real work lives. npm install -g flowise followed by npx flowise start gives you a working builder at localhost:3000 in minutes — I had a flow answering questions over a PDF before my coffee cooled. But production scale means configuring persistence (a real database instead of the default), secrets management, the message queue and worker setup for horizontal scaling, and your own monitoring beyond the built-in observability. None of that is hidden — the docs cover it — but it is DevOps work that a hosted agent platform absorbs for you. Self-hosters should price their time honestly.

Third, the prediction metering on cloud is the number to watch. The free tier's 100 predictions per month is a taste; a flow under real use eats through that in an afternoon of testing, and Starter's 10,000 is the first tier that supports an actual internal tool. Because a single user-facing question can fan out into several model calls inside one flow (retrieval, rewrite, answer, evaluate), I found my prediction consumption running higher than naive per-message estimates suggested. The evaluations and metrics features helped me see where the calls were going.

Fourth, human-in-the-loop nodes work as advertised and are underused in the wild. Inserting a review step where an agent proposes an action and a human approves it changed the character of the automation: I used it for an agent that drafted outbound responses and held each one for approval before any tool call fired. That pattern — agents propose, humans release — is where I think this category belongs in production, and Flowise makes the wiring two nodes.

Where it falls short

The learning curve is real: the canvas exposes LangChain's conceptual surface (chains, agents, memory types, retrieval strategies), and a beginner who does not know what an embedding is will be lost in the node picker. Debugging large Agentflow canvases gets unwieldy — with dozens of nodes, the failure can sit three hops from the symptom, and the tooling for that (logs, traces) is competent rather than delightful. The cloud's free tier is close to decorative, and I would not judge Flowise's performance on it. Enterprise identity and permission features are license-gated, so the "free self-host" story frays exactly where larger teams need it. The ecosystem moves fast, which cuts both ways: node names and options shift between versions, community templates vary in quality, and model-provider changes land quickly. There is also no fully managed, opinionated production runtime — if you want someone else to own uptime, scaling, and failover, you are looking at the cloud tiers and their limits, or a competitor with a managed SLA.

A note on community and support: the project's GitHub activity is heavy and the docs cover the standard recipes (deploy, authenticate, scale), but there is no formal SLA behind the free tier — support is community channels and issues. The hosted Pro plan adds priority support and admin roles, which is where small teams running an internal tool on Flowise Cloud should realistically land, and the first-month-free offer on Starter makes that an easy experiment.

Verdict

Flowise gives builders the control of code-based LLM development through a canvas you can actually reason about — retrieval, multi-agent flows, human-in-the-loop review, and deployment from the same project. I rate it 4.0: depth of core functionality is high for visual LLM orchestration and the self-host economics are attractive, while usability is aimed at people who already understand the concepts and the license split plus cloud tiering complicate the "free" story. For developers and technical teams prototyping and shipping LLM applications, it is one of the first tools I would try; for non-technical users or teams wanting a fully managed runtime, it is more machinery than you need.

Sources

• https://flowiseai.com/
• https://github.com/FlowiseAI/Flowise
• https://github.com/flowiseai/flowise/blob/main/LICENSE.md
• https://docs.flowiseai.com/getting-started
• https://www.npmjs.com/package/flowise
• https://lindy.ai/blog/flowise-pricing
• https://toolradar.com/tools/flowise/pricing
• https://rottenwifi.com/services/flowise
• https://itechguides.com/products/flowise/pricing
• https://jacar.es/en/flowise-low-code-self-hosted-agents
• https://selfhost.directory/project/flowise-ai
• https://buildfastwithai.com/ai-tools/flowise