AgentFlux is a self-hosted visual workflow builder for AI agents

An agent platform that keeps the documents, the model and the audit trail on one computer is a rare thing. AgentFlux is a new open-source project that does this. The GitHub repository has 171 stars as of 29 October. The repository was created on 27 October by developer Dave Messew. The first public release, v0.2.0, is from the same day.

The project is a visual workflow builder with an AI agent built in. The user drags preconfigured steps onto a canvas, connects them, and runs the workflow. The agent can also build the workflow itself from a task description. Everything runs inside one Docker container with a browser interface at port 4310.

What v0.2.0 includes

The README names 17 preconfigured steps and eight templates. Workflows can branch, undo and redo, and import or export as JSON. The AI agents feature lets a user describe a task and pick data sources and a model. The agent then starts manually or on a schedule.

A separate feature called "Analyse video" turns a screen recording into an editable workflow. The agent reads up to 24 evenly spaced JPEG frames from the recording. It asks short questions if a step is unclear and continues from the user's answers. The README caps the input at 250 MB and 30 minutes. This feature is in the source code but not in the prebuilt v0.2.0 Docker image.

The data layer is SQLite. Imports include PDF, DOCX, XLSX, CSV, JSON and text. Models are user-supplied: Ollama, LM Studio, vLLM and any OpenAI-compatible endpoint. The release includes the app, the built interface and Node.js. The user does not install Node or pnpm separately.

What the agent can and cannot do

The documentation is specific about limits. Generated workflows contain only executable steps for existing connections. The analysis cannot control other desktop applications. If a suitable connection is missing, the agent asks about the target and names an available alternative.

Cloud model calls and external HTTP POST steps need two separate approvals: one in Settings and one in the workflow. HTTP GET requests can send information in URLs and headers. An optional redaction step uses rules and may miss sensitive information. The README states these conditions in the "Operations and data" section. A user reads them before pointing the agent at company documents.

Workflows, documents, credentials and run logs live in a named Docker volume called agentflux-data. Replacing the container does not delete them. A backup must include the entire volume, including the .vault-key file that decrypts stored credentials.

What this is not

The AGPL-free Apache 2.0 licence allows commercial use. But v0.2.0 is a pre-1.0 release. The prebuilt image is linux/amd64 only. On Windows, the user runs Docker Desktop in Linux container mode. Other architectures need a build from source.

The configuration binds port 4310 to localhost and comes with no login. The README states the design plainly: AgentFlux has no login and is designed for a single user. A team that shares one instance has to add its own access-control layer. The network also has to be set up for it.

There is no benchmark against the hosted platforms. The README does not claim higher accuracy or lower latency than cloud agents. It names the trade: a user owns the pieces and runs them locally, in exchange for pre-1.0 maturity and single-user design.

For a team weighing a cloud agent contract against a local alternative, AgentFlux is a cheap test. It takes a Docker install and an existing model server. The run logs show what the agent did, in the user's own volume, on the user's own machine.

Source

• davemessew/AgentFlux, GitHub
• AgentFlux 0.2.0 release, GitHub