Drop Footage in a Folder and Chat to Edit: 'video-use', an Open-Source Agentic Video Editor for Claude Code

On September 24, 2026, the open-source community unveiled 'video-use', an agentic video editing tool that enables creators to drop raw video footage into a folder and converse with Claude Code in the terminal to produce a polished final video (final.mp4). Bypassing conventional non-linear editor (NLE) timelines and complex inspector menus, the agent reads audio and visual cues to orchestrate trimming, audio cleanup, color grading, custom subtitling, and motion graphics overlays through a unified command-line workflow.

Image source: https://github.com/browser-use/video-use

Traditional video post-production has long demanded hours of manual trimming inside heavyweight NLE suites such as Premiere Pro or DaVinci Resolve—sorting through dozens of retakes, scrubbing playheads, slicing false starts, and aligning audio fades by hand. While recent AI video editing services offer automated cuts, most operate as closed cloud platforms constrained by fixed templates and recurring subscriptions. In contrast, video-use integrates directly into Claude Code as a local, code-driven agentic skill, executing transparent FFmpeg and Python workflows over the creator's local files.

Dual-Layer Perception Architecture: Token-Efficient Precision Cutting

Deploying large language models to video editing presents a core architectural bottleneck: ingesting raw, uncompressed video frames rapidly exhausts context windows and computational budgets. Feeding minutes of high-resolution video directly into multimodal vision models remains prohibitively slow and expensive.

video-use circumvents this challenge through a Dual-Layer Perception architecture, ensuring the language model never has to ingest full-length video streams directly.

• Layer 1: Always-Loaded Audio Transcript: The tool initiates a single ElevenLabs Scribe API call per source video to extract word-level timestamps, speaker diarization, and non-verbal acoustic events such as laughter, sighs, and applause ((laughter), (applause), (sigh)). The resulting metadata across multiple takes compresses into a lightweight ~12KB markdown file (takes_packed.md). This transcript serves as the primary script representation that the LLM reads continuously, enabling word-boundary cut calculations with negligible token overhead.
• Layer 2: On-Demand Visual Composite: Visual data is summoned strictly when necessary. During ambiguous pauses, facial expression comparisons across alternative takes, or cut-boundary sanity checks, the agent invokes timeline_view. This utility compiles a composite PNG featuring a multi-frame filmstrip thumbnail strip, an audio waveform graph, and synchronized word labels across the requested time window, providing precise visual grounding only at critical decision points.

Automated Pipeline: From Filler Removal to Color Grading and Dynamic Motion Graphics

video-use supports diverse video formats—including talking heads, interviews, tutorials, travel vlogs, and montages—delivering end-to-end automated post-production:

1. Intelligent Trimming and Clean Transitions: The engine scans for verbal disfluencies ('umm', 'uh'), false sentence starts, and dead space between takes, drafting and applying a refined cut list. Every audio splice receives an automated 30ms audio fade, eliminating the popping and clicking artifacts common in programmatic audio edits.
2. Segment-Level Automated Color Grading: Depending on scene mood or aesthetic goals, the tool applies preset color corrections—such as 'warm cinematic' or 'neutral punch'—or constructs customized FFmpeg color filter chains across designated segments.
3. Custom Subtitle Burn-In: To enhance retention and accessibility, video-use burns hardcoded subtitles into the export by default using a rapid 2-word UPPERCASE display cadence, with full user customization available for typography, alignment, and sizing.
4. Parallel Sub-Agent Motion Overlays: Beyond static text, the pipeline generates dynamic motion graphics and visual explainers using Manim, Remotion, and PIL. Dedicated sub-agents spin up concurrently to write, render, and composite each graphic asset directly onto the primary timeline at specified timecodes.

Self-Evaluating Quality Gates and Session Continuity

To prevent hallucinated cuts or disjointed edits, video-use implements an internal self-evaluation loop prior to presenting outputs to the creator.

Before final delivery, the agent evaluates rendered cut boundaries, verifies frame continuity across transitions, and audits audio smoothness. If a transition feels jarring or an edit clips a syllable, the system adjusts cut parameters and re-renders the segment automatically.

The project also features robust session and workspace engineering:

• State Persistence via project.md: Editing decisions, project parameters, and cut lists persist in a local project.md file. Creators can pause work and resume days later without losing conversational context or edit history.
• Clean Workspace Isolation: All temporary renders, cache files, and exported masters route cleanly to <videos_dir>/edit/, ensuring the root skill directory remains pristine.
• MIT-Licensed Open Source: Distributed under the permissive MIT license, video-use gives creators and developers full freedom to inspect, customize, and extend the pipeline locally.

Sources

• GitHub Repository: browser-use/video-use
• Official Announcement on X (@RoundtableSpace): video-use Release Announcement and Demo Thread