I Built a Local AI Podcast Editor Because I'm Done Renting My Own Workflow

The frustration behind this project: with most content tools, your files live on their servers, your workflow lives in their interface, and your access lives on their billing page. DarkRoom is my answer to that, a local-first multi-camera podcast editor that takes raw footage and produces full edits, vertical cuts, and social shorts without uploading anything.

The stack is deliberately plain and swappable. Python and Flask for the backend, local OpenAI Whisper for transcription, the Claude API for text-only analysis, FFmpeg for rendering, vanilla HTML/JS/CSS for the UI, and plain JSON files in a local folder for storage. The flow: drop in 2 to 4 synced camera files tagged by speaker, Whisper transcribes locally with speaker attribution, Claude reads the transcript and emits a JSON edit decision list, you review and tweak those decisions in the browser, and FFmpeg renders the outputs. Core processing has no external dependencies, API cost is around ten cents an episode, and because the LLM step is replaceable you can run it fully offline with a local model.

Key takeaways

• Own the artifacts: the edit decision list is readable, git-friendly JSON, not a proprietary project file
• Local-first isn't primitive: Whisper plus FFmpeg plus a thin AI layer covers transcription, speaker tagging, and cut planning on a normal laptop
• Keep the AI step optional: swapping Claude for a local LLM is a drop-in change that keeps the pipeline offline-capable

Who this is for

Podcasters and video creators tired of subscription tools and cloud lock-in, and developers interested in stitching Whisper, an LLM, and FFmpeg into a real workflow. It's a proof of concept, not a polished app, so expect to be comfortable running a Flask server locally.

The full write-up, with the architecture table and the roadmap, is on DEV.to. Code is at github.com/chaotictoejam/darkroom.