The standard advice for AI-assisted video production involves a stack of subscription tools: a script generator, a voiceover service, a stock footage platform, a video editor with AI features. The monthly cost adds up fast and most of the tools are wrapping the same underlying models you could run yourself. Building a local video scripting pipeline from free and self-hostable tools is not only possible, it produces a workflow you actually control rather than one that breaks every time a SaaS product changes its pricing or API.
The goal here is a complete pipeline: topic input goes in, a script comes out, a voiceover is generated from the script, footage is sourced, and the final video is assembled. Every step runs locally or through a free API tier. No subscriptions. No per-minute voiceover charges. No footage licensing fees per clip. This is the architecture behind ReelCast-style local video automation and the same principles that make n8n content syndication pipelines work without recurring tool costs.
Script Generation with a Local LLM
Ollama running a quantized 7B or 8B model handles script generation well for most video formats. The key is prompt engineering for the specific output format you need. A voiceover script has different requirements from a talking-head script or a documentary narration, sentence length, pacing markers, reading level, and information density all need to be specified explicitly rather than left to the model’s defaults.
Structuring your script prompt around the video’s specific requirements produces better output than a generic “write a script about X” approach. Define the target duration, the intended audience, the tone, the key points that must appear, and the format of any on-screen text cues. A model that knows it’s writing a 90-second explainer for a technical audience with specific terminology requirements will produce a more usable first draft than one handed a vague topic.
Voiceover Without a Subscription
Edge TTS is Microsoft’s text-to-speech engine exposed through an unofficial Python library that calls the same API used by Microsoft Edge’s read-aloud feature. The voice quality is well above what most people expect from free TTS and it supports dozens of language and accent variants. It runs as a Python command-line tool, integrates easily into automation scripts, and produces MP3 output suitable for video production without any account requirements.
Kokoro TTS is a local alternative worth testing if you want complete offline capability. It’s a smaller model than the large commercial TTS systems but produces natural-sounding output for common English speech patterns. The quality is acceptable for narration use cases where the voice is background to visual content, though it’s not at the level of a professional voice actor for hero content. Both tools can be integrated into an n8n workflow with Ollama as processing nodes without requiring any paid API access.
Footage Sourcing via Pexels API
Pexels provides a free API with generous rate limits that returns royalty-free video clips searchable by keyword. The API response includes direct download URLs for clips in multiple resolutions, which means you can automate footage sourcing as part of your pipeline rather than manually browsing a stock library. Extract keywords from your script, query the Pexels API for relevant clips, download the best matches based on duration and resolution criteria, and pass them to your assembly step.
The practical limitation is that keyword-to-footage matching is inexact. Searching for “machine learning workflow” returns office and technology generic footage rather than anything specific to your topic. Build your footage queries around visual concepts rather than technical terminology, and plan for a manual review step before final assembly for anything that will be published publicly.
Assembly with ffmpeg
ffmpeg handles the entire assembly step: concatenating footage clips, layering the voiceover audio track, adding title cards and lower thirds as image overlays, normalizing audio levels, and exporting the final video in any format you need. It’s command-line only, which means it integrates cleanly into automation pipelines without a GUI dependency.
The ffmpeg command for a basic assembly, background video with voiceover audio mixed at specified levels, fading in and out — is around ten lines. For more complex productions with multiple clips timed to audio cues, a Python script that generates ffmpeg commands programmatically is more maintainable than a single complex command. The assembled pipeline from script to final video can run end-to-end without any tool that charges per minute, per clip, or per seat.
A fully local video pipeline doesn’t replace a professional production workflow. It replaces the expensive SaaS version of an automated workflow that was never going to produce broadcast-quality output anyway. If you’re producing informational or educational video content at volume, the economics of free tools plus your own compute make more sense than subscriptions that add up to hundreds of dollars a month for the same functional result.
