FireRedOpenStoryline
title: FireRed-OpenStoryline type: tool created: 2026-06-22 last_updated: 2026-07-20 related: ["radar/tools/EditMind", "radar/tools/Yapsnap", "radar/tools/Sentrysearch", "radar/techniques/SelfImprovingAgentHarness", "radar/techniques/BuildingEffectiveAgents"] sources: ["https://github.com/FireRedTeam/FireRed-OpenStoryline"] radar_quadrant: Tools radar_ring: retired radar_position: outer
FireRed-OpenStoryline is an open-source LLM-powered video editing agent that translates natural language instructions into editing operations via an MCP (Model Context Protocol) server backbone. The agent interprets user intent, plans a sequence of steps, and orchestrates MoviePy and FFmpeg tool calls to produce edited video output. Built on LangChain with a FastAPI/Uvicorn web interface. Requires Python 3.11+ and Conda; a Docker image is available. Licensed Apache 2.0. The repository had approximately 3,000 stars and 347 forks as of June 2026, with active development since early 2026.
The system centers on an MCP server (src/open_storyline/mcp/) that exposes editing tools the LLM planner calls during task execution. Three layers operate together: an LLM planning layer that interprets conversational directives, video processing nodes (src/open_storyline/nodes/) that wrap MoviePy and FFmpeg operations, and an agent skills library (src/open_storyline/skills/) that stores reusable workflows. Agent memory and state persist in src/open_storyline/storage/. The entry point is agent.py.
This constitutes a genuine agent loop: the LLM plans, calls tools, receives results, and refines -- rather than executing a fixed pipeline.
- Clip cutting, swapping, and resequencing via natural language
- Script editing with control over tone, font, color, stroke, and positioning
- AI-generated transitions derived from ending and opening frames plus a natural language description
- Beat-synced background music placement with mood-based auto-recommendation
- ASR-based rough cut (added March 2026): automatic removal of filler words, disfluencies, and repeated sentences, with timestamp-aligned segmentation using local
torchaudioinference
Complete workflows can be saved as named Skills. Swapping in new media and invoking a Skill replicates the full editing style without re-specifying instructions. Skills are compatible with Claude Code, OpenClaw, and experimentally with Codex, making them composable within broader agent systems.
Eight demo video styles are showcased: Zhongcao Style, Humorous Style, Product Picks, Artistic Style, Unboxing, Talking Pet, Travel Vlog, and Year-in-Review. Live demos run on HuggingFace and ModelScope. Demos in the second row use the Xiaohongshu "AI Clip" asset library, indicating real-world platform adjacency. The project is featured on HelloGitHub. As of June 2026 the repository had 17 open issues, 7 open pull requests, and 50 commits.
No face detection or object detection features are present. Time-lapse is not documented as a supported operation. Direct timestamp-based navigation by user input is not described; timestamp handling is ASR-driven rather than query-driven.
Retired from Tools as disqualified (2026-07-20).
FireRed-OpenStoryline entered at Assess/inner on the strength of its agent-loop architecture, reusable Skills, Claude Code compatibility, and working demos. A first-person trial in the radar/techniques/PersonalVideoProcessingPipeline (the pynanoclip project) disqualified it for that use case. The only available hand-off from structured discovery data (identity plus timestamps) is the natural-language prompt, and the agent routes such a prompt into its speech_rough_cut node, producing an empty timeline; render_video then fails with ValueError: timeline result has no video track, the agent retries once and recommends manual editing. No prompt phrasing was found that makes the hand-off work. A pre-cut workaround, feeding it a plain clip instead, rendered but scored only 2 of 3 on a pre-agreed judgment call and required local vendor patches that do not survive a git pull.
The deterministic recipe-to-CapCut approach (radar/tools/CapcutCli, sibling playcapcut project) replaced it: highlight windows and timelapse bridges compile from a versioned JSON recipe into a CapCut draft, with no agent in the loop. Disqualified marks FireRed-OpenStoryline as failed-in-practice for turning discovered footage into a finished edit; do not start new work against it for this purpose. It remains distinct from radar/tools/EditMind (video search, no agent loop) and radar/tools/Yapsnap (transcription primitive), but neither those distinctions nor its demos survived contact with a real hand-off.
