AgentFlux 0.3.0

Recurring tasks now reuse a saved plan, ask for missing details in text and deliver reports to a local folder. New or changed inputs are processed after successful delivery; repeat runs reuse existing report batches.

• Guided task setup, including data preparation without a model.
• XLSX sheet/header preview, named fields and source coordinates.
• Structured task reports with source references and unresolved owners or deadlines.
• Segmented video analysis with screen-change sampling and local speech recognition.
• Progress, cancellation, explicit external-processing consent and cleanup of raw media.
• Modular frontend with shared types and TypeScript checks in the production build.
• Updated blue interface, English documentation and screenshots.

Install

Download the Docker archive and compose.yaml into one folder:

docker load -i agentflux-0.3.0-linux-amd64.tar.gz
docker compose up -d --wait

Open http://localhost:4310. The archive contains both Linux/amd64 images: AgentFlux and its local speech service. FFmpeg and the Whisper base speech model are included. Connect your own language/vision model for AI tasks and video interpretation.

For upgrades, keep the same Compose project and data volume, including .vault-key. Use the new Compose file and retain any custom mounts. INSTALL.md includes the full instructions; verify downloads against SHA256SUMS.txt.

Video uploads support up to 250 MB and 30 minutes within the selected analysis budget. Generated workflows use supported integrations; they do not replay arbitrary desktop actions.

Validation

Production build and 10 focused checks passed, along with Docker startup, report delivery and an upgrade from v0.2.0 retaining documents and encrypted credentials. Media checks used actual FFmpeg extraction and local transcription. Vision-model responses were a protocol fixture, so recognition quality across real vision models has not been benchmarked.