davemessew/AgentFlux

AgentFlux

Your data. Your models. Your workflows.

Agentic AI platform for visual workflows and autonomous agents.
Import files, automate tasks and keep results local.

Download the Docker release · Website · Documentation · Report an issue

What you can do

AgentFlux connects documents, data sources and language models in a visual workflow. Build each step yourself or let an agent create a workflow from a task description. The application runs on your computer or server, with a browser interface.

| | In AgentFlux | | :--- | :--- | | Visual workflows | 17 preconfigured steps, eight templates, branching, undo/redo and JSON import/export. | | AI agents | Describe tasks, select sources and a model, then start manually or set a schedule. | | Video to workflow | Identify actions in a recording, review them with timestamps and turn them into an editable workflow. | | Knowledge and learnings | Store documents locally, review observations from runs and reuse instructions as skills. | | Your own models | Connect Ollama, LM Studio, vLLM and OpenAI-compatible endpoints. | | Files and connections | Import PDF, DOCX, XLSX, CSV, JSON and text; connect folders, HTTP endpoints and approved programs. | | Traceable runs | Inspect steps and results, cancel running tasks and export results as JSON or CSV. |

Create a workflow from a video

Open Agents → Analyse video, select a recording and a model connection with vision support. If details are missing or a step is unclear, the agent asks short questions. Reply in text and it continues the analysis using your answers. Once the workflow is fully mapped, you can review the steps with timestamps. Choose Open in editor to save and run the workflow.

The recording stays in your browser. Up to 24 evenly spaced JPEG frames are sent to the selected model server for analysis. Audio is not analysed, and brief actions between frames may be missed. Videos up to 250 MB and 30 minutes are supported if your browser can read the codec. Cloud models require explicit permission to receive unredacted images.

Generated workflows contain only executable steps for existing connections. The analysis does not control other desktop applications. If a suitable connection is missing, the agent asks about the target and an available alternative. Video analysis is included in the current source code; the prebuilt v0.2.0 Docker image does not include it yet. Use a build from source for this feature.

Get started with Docker in minutes

You need Docker with Compose. The release includes the application, built interface and Node.js. You do not need to install Node or pnpm locally.

Use the prebuilt image

Download these two files from release v0.2.0 into the same folder:

• agentflux-0.2.0-linux-amd64.tar.gz: the prebuilt Docker image.
• compose.yaml: the startup configuration.

~~~sh docker load -i agentflux-0.2.0-linux-amd64.tar.gz docker compose up -d --wait ~~~

Open localhost:4310. Done.

The prebuilt image targets Linux/amd64. On Windows, use Docker Desktop in Linux container mode. You can build from source for other architectures; they were not tested for this release.

A look inside

The screenshots show the current source version. The published v0.2.0 Docker release still uses the green colour scheme; the new design is included when you build from source.

Build workflows on the canvas

Connect sources and processing steps, configure each step and follow its execution.

Agents Tasks, models and recurring jobs in one place.	Knowledge base Search documents and processed content locally.

Your first workflow

1. Open Templates and choose Clean and save data. This template uses synthetic sample data and does not need an AI model. 2. Start the workflow and open its run results. 3. Connect your model server under Models to summarise documents or use agents. 4. Create a task under Agents and select the allowed sources and model.

The Docker image does not include a model. Use an existing local server or an external provider with your own credentials.

| Model server on your computer | Base URL from the container | | :--- | :--- | | Ollama | http://host.docker.internal:11434 | | LM Studio / OpenAI-compatible | http://host.docker.internal:1234/v1 | | vLLM | http://host.docker.internal:8000/v1 |

Adjust the port and model ID for your server. The model server must accept connections from the container.

Operations and data

~~~sh docker compose ps # Status and healthcheck docker compose logs -f # View server output docker compose stop # Stop docker compose up -d --wait # Start again ~~~

Workflows, documents, credentials and run logs stay in the named agentflux-data volume. Replacing the container does not delete them. Back up the entire volume including .vault-key; docker compose down -v deletes the volume.

[!IMPORTANT]

The supplied configuration binds port 4310 to localhost. AgentFlux has no login and is designed for a single user. Shared or public access requires an access-control layer and suitable network configuration.

Cloud model calls and external HTTP POST steps require approval in Settings and in the workflow. HTTP GET requests can send information in URLs and headers. Optional redaction uses rules and may miss sensitive information. See the installation guide and operations guide.

Documentation and development

| Start here | Contents | | :--- | :--- | | Docker release | Prebuilt image, Compose file, source code and SHA-256 checksums | | Install the release | Download, verify, start and update | | Installation and models | Docker, native installation and model connections | | Workflows and agents | Editor, steps, templates, learnings and skills | | Operations and configuration | Environment variables, import folders, backups and programs |

~~~text src/ Web interface and visual editor server/ API, agents, execution and SQLite storage shared/ Workflow rules and templates examples/ Sample local program export scripts/ Docker release checks docs/ Usage, operations and screenshots ~~~

For local development: Node.js 24+, pnpm 11.25.0, pnpm install --frozen-lockfile, then pnpm dev.

Run the Docker check after building:

~~~sh docker build -t agentflux:0.2.0 . node scripts/smoke-docker.mjs agentflux:0.2.0 ~~~

The check uses its own container and volume. It covers the API, built interface, an example run and data persistence when replacing the container.

Licence

Apache License 2.0. Bundled dependencies and fonts retain their respective licences.