GGitHub采集材料
VibeFrame:给视频生成设定费用上限
发布时间未知 · 北京时间
原文与观测指标
1 个来源GGitHub
查看原文 VibeFrame:给视频生成设定费用上限
vericontext星标182采集时的累计值
每小时增长待观察星标 / 小时 · 观测区间平均值
最近提交10-28 08:25提交时间不是项目发布时间
发布时间:未知 · 首次发现 10-28 08:32 · 北京时间
指标采集:10-28 08:32 · 指标已取得
项目文档有线索安装说明有线索使用说明有线索示例有线索
仅反映文档与示例线索,未实际运行项目。
摘要
VibeFrame 通过 CLI 和 MCP 让编程智能体根据文字简报生成视频。可先用 dry run 估算各阶段费用,再用 --max-cost 设置硬性上限;超预算时会拒绝构建并返回可调整方案。
摘要由 AI 生成,请以原文为准。观测历史 · 最近 1 条 展开 / 收起
缺失不计为零;增长只比较同一指标来源的有效记录,按实际间隔折算每小时。上游缓存可能延迟数值更新。
| 采集时间 · 北京时间 | 指标 | 数值 | 来源 | 状态 |
|---|---|---|---|---|
| 10-28 08:32:16 | Star | 182 | GitHub 公开页面 | 已取得 |
采集原文 展开 / 收起
vibeframe
Frontier AI video generation for coding agents - Seedance, Runway, Veo, Kling on your own keys, behind a hard cost cap. CLI + MCP.
项目说明 展开 / 收起
VibeFrame Let your coding agent generate real video, on your own provider keys, under a spend ceiling it cannot cross. VibeFrame is a CLI and MCP server for Claude Code, Codex, Cursor, or any bash-capable agent. It turns a written brief into a plan, generates the assets from frontier models (Seedance, Runway, Kling, Gemini Omni), and renders a finished MP4. Every paid step sits behind a dry run and a hard --max-cost ceiling, and every failure comes back as machine-readable recovery actions instead of a stack trace. One photographer across a single arctic night, 1080p, generated end to end. ▶ Watch the full render Know the price before anything is spent --dry-run prices the whole build without calling a provider or needing a single key. Pair it with --max-cost and the build refuses to start when the estimate is over your ceiling: vibe build film --dry-run --max-cost 3 --json { "success": false, "error": "Estimated cost $10.93 exceeds --max-cost $3.00.", "code": "COST_CAP_EXCEEDED", "retryWith": [ "vibe build . --stage all --skip-backdrop --json", "vibe build . --stage all --max-cost 10.93 --json" ], "data": { "plan": { "estimatedCostUsd": 10.93, "...": "the full priced plan" } } } The envelope goes to stderr and the command exits 1, so an agent loop stops instead of guessing. retryWith gives it the two cheaper ways forward it can take without asking you, and data.plan carries the priced plan the refusal was based on, so it never has to re-run just to learn the estimate. The same dry run names the keys each stage would need, so you find out what a video costs before you own an account: Keyframe generation will need openai (OPENAI_API_KEY), but no key/config is available. Video generation will need seedance (FAL_API_KEY), but no key/config is available. Your keys, your bill, your ceiling. Where VibeFrame fits It wraps lower-level composition engines rather than replacing them: Layer Owns Remotion React-based programmatic video and component-driven motion graphics. Hyperframes HTML/CSS/JS scene composition and deterministic browser capture. VibeFrame Frontier-model generation on your own keys, dry runs and the hard --max-cost ceiling, storyboard/design files, build reports, render inspection, edit/remix commands, and host-agent guidance. Read that table as one sentence: scene composition is delegated, generation is not. If the job is only HTML scene authoring and rendering, use Hyperframes directly - it is the better tool for that, and vibe scene install-skill installs its skill for you either way. Reach for VibeFrame when the job involves paid generation: frontier image and video models, character continuity across scenes, narration, cost ceilings, build reports, or editing steps around the composition layer. One consequence worth knowing before you install: the whole pipeline runs locally for $0 without a provider key, because that free half is the delegated half - Kokoro narration, agent-authored HTML/CSS scenes, headless Chrome plus FFmpeg. It is the cheapest way to see whether the workflow fits you, and the full walkthrough is in docs/hyperframes.md. It is not the point of the tool: the paid generation path below is. VibeFrame is not affiliated with HeyGen. See CREDITS.md for dependency and provenance notes. Install curl -fsSL https://vibeframe.ai/install.sh | bash vibe doctor Needs Node.js 20+, FFmpeg, and Chrome (or Chromium). vibe doctor verifies all three and lists which commands each provider key unlocks. Generation is the only thing that needs a key. FFmpeg edits, Kokoro narration, scene composition, detection, timeline work, and rendering are all local. Add OPENAI_API_KEY, FAL_API_KEY, GOOGLE_API_KEY, RUNWAY_API_SECRET, or KLING_API_KEY for the providers you actually use, via vibe setup or the environment. Package names, config locations, and building from source The CLI is published as @vibeframe/cli (binary vibe) and the MCP server as @vibeframe/mcp-server. There is no bare vibeframe npm package from this project - that name belongs to an unrelated package, so npx vibeframe will not run this tool. The installer places the CLI under the XDG data directory (~/.local/share/vibeframe by default). User-scope keys live in ~/.vibeframe/config.yaml; vibe setup --scope project writes ./.vibeframe/config.yaml. When a project config exists at or above your current directory, VibeFrame uses it in isolation and does not merge user-scope keys. git clone https://github.com/vericontext/vibeframe.git cd vibeframe && pnpm install && pnpm build pnpm vibe --help Build a video from a brief vibe init launch --from brief.md # or --from "30-second launch video for VibeFrame" vibe build launch --dry-run --max-cost 5 # price it vibe build launch --max-cost 5 # generate vibe render launch -o renders/final.mp4 # render vibe inspect render launch --cheap # check the result brief.md can be messy notes, links, or a single line. What init produces is a project you then own: launch/ STORYBOARD.md title, duration, aspect, providers, cast - and the direction prose DESIGN.md palette, typography, layout, motion, transitions scenes/ 01-hook.md one scene per file, in play order 02-proof.md 03-close.md Each scene is a markdown file. Its frontmatter carries the cues, its body carries the direction: --- type: Scene duration: 5 narration: "Start with a storyboard. VibeFrame turns each beat into a render plan." backdrop: "Clean developer terminal beside structured storyboard cues" video: "media/broll.mp4" voice: "alloy" --- Open on the terminal. Let the storyboard cues land before the camera moves. A cue's value is either a prompt to generate from, or a project-relative path to a file you already have - video: "media/broll.mp4" above reuses your footage instead of generating any. Scenes play in filename order, so vibe storyboard move reorders by renumbering. Edit the files directly, or ask a coding agent to research and revise them. The rest is convention: media/ for footage you supply, assets/ for generated ones, render