Floneum – Visual AI Workflows with WASM Plugins

TL;DR

TL;DR: Floneum is a graph-based visual editor for building AI workflows. It runs local LLMs with no external dependencies, chains community WASM plugins in a sandboxed environment, and supports plugin development in Rust, C, Java, or Go.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: floneum.com
• Source repository: github.com/floneum/floneum — verified README
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• HN launch thread: news.ycombinator.com/item?id=42228166 (61 points)
• Docs: floneum.com/docs — verified installation, features
• Current main branch — GitHub API confirms pushed_at: 2026-08-08

What Is Floneum?

Floneum is a graph-based visual editor for building AI workflows. According to the official docs:

“Floneum is a graph editor for AI workflows with a focus on community made plugins, local AI and safety.”

It has two sides:

• User side: A drag-and-drop graph editor anyone can use to chain AI tasks without writing code.
• Developer side: A WASM-based plugin system that lets you write plugins in Rust, C, Java, or Go and publish them to the community.

The project also includes Kalosm, a Rust crate ecosystem for running pre-trained language, audio, and image models locally.

Key Features

From the verified docs and GitHub README:

• Visual graph editor — chain community plugins with local AI models in a drag-and-drop interface, no programming knowledge required.
• Local LLM execution — uses the llm Rust library to run large language models with no external dependencies or GPU required.
• WASM plugin sandbox — all plugins run isolated in a WebAssembly sandbox. Plugins only access what the environment allows, removing the need to fully trust third-party code.
• Multi-language plugins — write plugins in any language that compiles to WASM. Official ergonomic Rust wrapper with macros available.
• Controlled text generation — plugins can constrain model output to structured formats (JSON, XML, etc.) using a technique similar to JSONFormer or Guidance.
• Built-in plugin library — includes 41 pre-built plugins for tasks like embedding, text generation, database operations, file I/O, browser automation, and RSS parsing.

Installation

From the docs, install via the releases page at github.com/floneum/floneum/releases.

No runtime dependencies required. GPU is optional but supported via CUDA for acceleration.

Deeper Analysis

Local-First Architecture

Floneum does not route data through any external API. The llm library handles local model inference, and the WASM sandbox contains plugins to their declared permissions. This makes it suitable for workflows that handle sensitive data.

Plugin Model

Plugins are WASM modules loaded at runtime. The Rust SDK exposes an ergonomic macro-based API:

#[floneum_plugin]
pub fn my_plugin() -> ... {
    // plugin logic
}

Plugins can request specific capabilities (file access, network, etc.) and users grant or deny them on load.

Kalosm: The Underlying ML Layer

The repo also contains Kalosm, which powers the model layer. Supported models include Llama (1B–70B, quantized), Mistral (7–13B), Phi-3 (2–4B), Whisper for audio transcription, Segment Anything for image segmentation, and Bert for embeddings. All can run on CPU, CUDA, or Apple Silicon (via WGPU/Fusor ML runtime).

Comparison to Similar Tools

| Feature | Floneum | LangFlow | Flowise | |---|---|---|---| | Local LLM | Yes (llm library) | Via API | Via API | | WASM sandboxing | Yes | No | No | | Multi-language plugins | Rust, C, Java, Go | Python only | JavaScript | | Built-in model runner | Yes | No | No | | Visual graph editor | Yes | Yes | Yes |

Practical Evaluation Checklist

• [ ] Download and install from GitHub releases
• [ ] Launch the graph editor and explore the plugin library
• [ ] Load a community plugin and add it to a workflow
• [ ] Connect a local LLM node and run a simple text generation task
• [ ] Try the structured generation feature to constrain output format
• [ ] Explore writing a simple plugin using the Rust SDK

Security Notes

The WASM sandbox restricts plugins to their declared permissions. Files, network, and other sensitive resources require explicit capability grants. Because plugins cannot escape their sandbox, even malicious third-party plugins cannot escalate privileges beyond what the user has approved.

No data leaves the local machine by default — all inference runs on-device.

FAQ

Q: Do I need a GPU to run Floneum? A: No. Floneum runs on CPU by default. GPU acceleration via CUDA is available when a compatible GPU and CUDA drivers are present.

Q: Which languages support plugin development? A: Any language that compiles to WebAssembly. The project provides official ergonomic wrappers for Rust. C, Java, and Go are also directly supported.

Q: How does Floneum compare to Ollama for running local models? A: Ollama manages model downloads and serves an API. Floneum focuses on workflow orchestration on top of local models — it uses the llm library directly and exposes a graph-based UI for chaining model calls with plugins, rather than an HTTP API.

Q: Is there a plugin marketplace? A: Community plugins are available at the project’s GitHub and shared via the plugin discovery system built into the editor. The docs mention 41 built-in plugins covering text generation, search, file operations, and browser automation.

Conclusion

Floneum fills a gap between pure local model runners (Ollama, llm) and full IDE-style workflow tools. Its visual graph editor makes workflow design accessible to non-programmers, while the WASM plugin sandbox keeps the system safe as you extend it with community contributions. Multi-language plugin support (Rust, C, Java, Go) is unusual and useful for developers who want to integrate existing codebases without rewriting them in JavaScript.

If you want to run local AI workflows on a visual canvas with a genuine sandbox around third-party code, Floneum is worth a look.

Source last checked: 2026-08-15 (commit main, GitHub API verified)