2026-10-26 · America/Los_Angeles · 社区动态 · #20

typed tool calls only, no raw ffmpeg strings - and the agent can render a frame to verify. it's rough, tell me what breaks

disclosure: i'm the author. i've been building kimo, a free, open-source (MIT) video editor with no GUI that AI agents drive over MCP - codex, cursor, opencode, or any other MCP client. the fresh angle here is that everything is typed tool calls only, no raw ffmpeg strings. and because it's built around the edit loop, the agent can actually render a frame to verify its own work before committing the timeline. why bother? raw ffmpeg is incredible but composing a timeline from shell commands is brutal for an agent - off-by-one frame errors, audio/video drift, guessing whether a filter chain…

热度 45.0 / 100；排名与评分保留该期记录。