davemessew/AgentFlux

AgentFlux is a self-hosted browser application for building LLM-driven workflows and agents visually. It suits a single user who wants documents, run logs and model calls to stay on their own machine or server, and who is willing to run an unauthenticated service on localhost.

What the workflow looks like

You import files (PDF, DOCX, XLSX, CSV, JSON, text) or connect folders, HTTP endpoints and approved programs, then connect them to processing steps on a canvas. The project lists 17 preconfigured steps, eight templates, branching, undo/redo and JSON import/export README. Each step can be configured and its execution followed; runs can be inspected, cancelled, and exported as JSON or CSV workflow docs.

Two mechanisms change how much work you do by hand. First, an agent can generate a workflow from a task description, and the generated workflow contains only executable steps for connections that already exist. Second, an experimental video-to-workflow path identifies actions in a screen recording, lets you review them with timestamps, and converts them into an editable workflow.

The knowledge base stores documents locally, and run observations can be reused as skills or instructions. Agents can be started manually or on a schedule.

Models and prerequisites

The Docker image ships no model; you point it at Ollama, LM Studio, vLLM or another OpenAI-compatible endpoint using your own credentials installation guide. The quickest path is the prebuilt v0.2.0 image plus Compose file: you need Docker with Compose, and the image targets Linux/amd64. Windows users can run Docker Desktop in Linux container mode. Other architectures require a source build and were not tested for that release README. Building from source needs Node.js 24+ and pnpm 11.25.0.

The prebuilt v0.2.0 image does not include video analysis or the newer interface shown in screenshots; both require building from source.

Single-user security boundary

The supplied configuration binds port 4310 to localhost only, and AgentFlux has no login; it is designed for one user. Shared or public access needs a separate access-control layer and network configuration. Cloud model calls and external HTTP POST steps require approval in Settings and in the workflow. GET requests can still leak data via URLs and headers, and optional redaction is rule-based and may miss sensitive information README, data and limits.

Video analysis caveats

The recording stays in the browser; up to 24 evenly spaced JPEG frames go to the chosen model server. Audio is ignored, and brief actions between frames may be missed. Videos up to 250 MB and 30 minutes are supported only if the browser can read the codec. Cloud models need explicit permission to receive unredacted images. Analysis cannot control other desktop applications, and if a needed connection is missing the agent asks about an alternative.

Data and licence

Workflows, documents, credentials and logs live in the agentflux-data volume and survive container replacement. Backups must include the whole volume and .vault-key; docker compose down -v deletes it operations guide. The code is Apache-2.0, with bundled dependencies and fonts under their own licences LICENSE.