Refly: an open-source Vibe Workflow platform that turns conversations into reusable processes

Refly calls itself the first open-source Vibe Workflow platform, and the point is not another chat box but a free canvas. You open several independent conversations on it, spread out ideas, sources and search results, connect them, and end up with a document or a repeatable process. It connects more than 13 mainstream large language models, parses multimodal files, ships a skill system and a knowledge base engine, and runs either in the cloud or self-hosted with Docker.

The trouble with a single thread is that it both forgets and blurs. By turn fourteen the framework you settled on, the options you rejected and the sources you cited all sit in one context, and the model starts losing the thread. Refly puts conversations on a canvas as nodes, each with its own context, so you decide exactly which nodes join this round and which stay parked. That is closer to how people actually work: several lines in motion, merged when needed.

The second layer is persistence. A one-off answer disappears; Refly tries to keep the process: where material came from, which skill processed it, which model ran, what came out. Being open source matters here, because neither the workflow nor the data has to be locked on a vendor's servers.

Main capabilities

Multi-threaded conversations

Several independent conversation contexts are managed in parallel and you switch freely between nodes, which is what makes a complex agentic workflow possible instead of cramming everything into one timeline.

Multi-model framework

More than 13 mainstream models are integrated with mixed scheduling and parallel processing, so switching models does not mean switching tools, and the chat interface stays uniform.

Multimodal handling

PDF, DOCX, TXT and more than seven file formats plus mainstream image formats such as PNG and JPG can be parsed in batches.

AI-driven skills

Capabilities including Perplexity AI and Stanford Storm provide web-wide search, knowledge retrieval, question rewriting and recommendations.

Context management and knowledge base

A temporary knowledge base can be built on demand and nodes chosen as context; the knowledge base ingests heterogeneous sources, runs semantic retrieval on RAG, and builds a knowledge graph.

Capture and clipping

One-click collection from platforms such as GitHub and Medium with automatic classification and tagging, plus clipping any web page to grow a personal library.

How to use it

1. Open the site, click start and sign in with Google or GitHub, or go straight to Refly Cloud.
2. Choose create canvas on the main screen and treat the piece of work as a workspace rather than a single question.
3. Open several conversations on the canvas and let different directions advance in parallel.
4. Revise with the AI's feedback and bring only the needed nodes into the next round.
5. Export to a document when done, or share it for collaborative review.
6. To self-host, deploy with Docker; the minimum is two CPU cores and 4 GB of memory. Private enterprise deployment is available on request.

Deployment options

• Cloud: open Refly Cloud and use every feature without configuration, a sensible first step before committing to self-hosting.
• Docker self-hosting: two CPU cores and 4 GB of RAM is the stated minimum, and a container command brings the service up.
• Local development: follow CONTRIBUTING.md in the repository, useful if you want to change code or wire in your own models.
• Enterprise: contact the team for a private deployment when data location is a hard requirement.
• Cost note: self-hosting removes a subscription but adds model API bills, key management and upgrade work, so run the monthly numbers.

Who it suits

• Researchers pushing a complex subject over weeks, keeping sources, hypotheses and conclusions in separate nodes that do not contaminate each other.
• People writing technical documentation or long articles, keeping ideation and the finished draft on one canvas.
• Anyone building a personal knowledge base, with capture, clipping and retrieval in a single line.
• Teams that care where data lives, since open source plus self-hosting keeps process and material on their own machines.
• People who use several models and want to pick per task instead of being tied to one.

Limits and caveats

• Threads and canvas carry a learning cost; anyone used to one long conversation will find node management harder than the question itself at first.
• More than 13 models means output quality depends on the model and the calling pattern you choose, and the same prompt behaves quite differently across them.
• The knowledge base relies on RAG and semantic retrieval, so chunking strategy and source quality decide retrieval quality; garbage in is still garbage out.
• Self-hosting means carrying model API costs, key management and upgrades yourself, and two cores with 4 GB is a floor rather than a comfortable setup.
• The source does not state quotas or billing for individual models, so the cost gap between cloud and self-hosted needs your own calculation.
• Retention policy and training use are not disclosed in the source; confirm before handling contracts or personal data.

Frequently asked questions

How is Refly different from an ordinary AI chat tool?

An ordinary tool is one line of conversation; Refly is a canvas where several conversations run in parallel, nodes can be selected precisely as context, and the output can settle into a reusable process.

Can it be self-hosted?

Yes. The project documents a Docker path with a stated minimum of two CPU cores and 4 GB of memory, and there are cloud and private enterprise options.

Which files are supported?

The source lists PDF, DOCX, TXT and more than seven document formats plus mainstream image formats such as PNG and JPG.

Do I need to write code?

No. The canvas is driven by conversation and node operations; only self-hosting and local development involve the command line and config files.