AI Skill Hub 推荐使用:梅隆多技能AI助手框架 是一款优质的AI工具。AI 综合评分 6.8 分,在同类工具中表现稳健。如果你正在寻找可靠的AI工具解决方案,这是一个值得深入了解的选择。
梅隆多技能AI助手框架 是一款基于 Shell 开发的开源工具,专注于 AI代理、工作流、自动化 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
梅隆多技能AI助手框架 是一款基于 Shell 开发的开源工具,专注于 AI代理、工作流、自动化 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 克隆仓库 git clone https://github.com/MelonS/MelonS-Agents cd MelonS-Agents # 查看安装说明 cat README.md # 按 README 完成环境依赖安装后即可使用
# 查看帮助 melons-agents --help # 基本运行 melons-agents [options] <input> # 详细使用说明请查阅文档 # https://github.com/MelonS/MelonS-Agents
# melons-agents 配置说明 # 查看配置选项 melons-agents --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量(覆盖配置文件) export MELONS_AGENTS_CONFIG="/path/to/config.yml"
A macOS multi-agent system driven by Claude Code. Latest tag is v0.4.0. Two production skills ship today; both are agentskills.io-spec compliant so they drop into Claude Code, Cursor, Goose, Gemini CLI, and the rest of the compatible-runtime set unmodified. Skill #1 —music-video. Music file in, 60-second 9:16 vertical short out. Per-genre color grade (six profiles) shapes generic Pexels B-roll into a genre-coded look; 23 ffmpeg shaders + phrase-aware structure (cuts onaubiotrackbeats, glitch micro-edits onaubioonsetdrum hits, restraint gated per preset) compose on top. The demo above is a noir-detective render; the grid further down shows six genre profiles side by side. Implementation underagents/missions/music-video/run.sh— the skill routes through the 5-agent mission pipeline (orchestrator + planner / resourcer / editor / qa) so re-rendering tuning flows into both surfaces. Skill #2 —job-hunt. Single seed keyword in, deduplicated markdown digest out.--seed "Problem Solver"expands to the 24-synonym role family companies actually use (FDE / Applied AI Engineer / Generalist / Founding Engineer / Forward Deployed / …) before fetching from 11 source plugins. Five are live-ready without an API key —global-ats(Greenhouse + Ashby + Lever boards across ~27 AI / SaaS companies including Anthropic / OpenAI / Cursor / Stripe / Notion / Datadog),global-remoteok,global-remotive,global-hn-whoshiring(HN monthly via Algolia),kr-worknet(정부 공공고용서비스). Two more (kr-wanted,kr-saramin) need an API key per source. End-to-end live test: 5,000+ raw postings filtered to ~200 matches. The skill is standalone (no orchestrator routing) —skills/job-hunt/scripts/run.shis the canonical implementation because every planner / qa stage would be near-empty for a mechanical filter / fetch / dedupe pipeline. Two ways to drive this repo. - Agent-driven (primary) — install Claude Code, point it at the cloned repo, type a mission. Claude Code edits files, commits, pushes. Cost: a Max subscription absorbs orchestration; the money firewall gates anything paid beyond that. - Script-only (fallback) —./scripts/bootstrap.shthen the bash scripts run standalone. No Claude Code needed; no commits or pushes happen automatically, but the rendered output is identical. Cost: $0 beyond the optional free Pexels API key. The scaffold is general-purpose. Short-form video is the v1 domain because the deliverable is visually verifiable and failure modes are fast to catch; the architecture itself doesn't assume short-form anything. New skills pick the shape that fits — see the Architecture section below for the three-shape model and what a future skill (movie / game / longform) would likely route through. Built on a single premise: automate the production pipeline, then let the system evolve its own logic. Every commit is one observable step of that evolution. The history is the agent system's own growth, not a record of its outputs (those stay local in gitignoredrecords/).
Engineering decisions, one page. docs/engineering-case-studies.md — nine production incidents and the minimum mechanism each one produced (Tier-1 routing, semaphore-bounded batch, content-quality feedback loop, three-layer reactive audit, shader-effects-in-ffmpeg / knowing-where-the-wall-is, onboarding-friction-kills-first-touch / zero-account demo path, declarative preset routing for genre-aware shaders, intervention-as-the-unmeasured-axis / autonomy signal + reduction levers, and the quality-bar-as-6-unenforced-contracts after the 2026-05-22 music-video QA pass). Each entry follows problem → constraint → decision → artifact.
- 23 ffmpeg shaders in scripts/music-video-shaders.sh across three stages — pond / halation / breathing / combo (first pass) + light_leak / duotone / vignette_pulse / scanline / chromatic_split / neon_edge / vhs / saturation_pulse / kaleidoscope / beat_burst / strobe / shake / color_burst / light_rays (genre-aware pass) + paper_grain / dust_speck / posterize / trail_echo / soft_bloom (Stage-2 + Stage-3). Cartoon / cel-shading deliberately deferred — see case study 5. - Genre-aware preset routing — 14-genre table in skills/music-video/data/genre-presets.yaml resolves a genre → preset → env overrides → post-shader chain (case study 7). Ambient / classical / dreamcore route to a separate scripts/music-video-stillzoom.sh (image + music → 60-second slow Ken-Burns) for genres where ANY cut violates the contract. - Per-genre base color grade — grade_profile field on every preset (kr_warm_pastel / hollywood_teal_orange / lofi_warm_grain / rnb_low_key / city_pop_neon / neutral) drives an ffmpeg curves + eq + colorbalance stage in scripts/music-video-grade.sh before shaders. Transforms generic Pexels B-roll into a genre-coded look before the effect layer. Research origin: docs/research/2026-05-22-music-video-pro-practices.md §2; visual A/B verdict in docs/research/2026-05-22-grade-profile-comparison.md. - Director-discipline shot plan (opt-in scaffold) — scripts/shot-plan.sh generates a per-segment intent layer from the lyric LRC + phrase boundaries before B-roll fetch, paralleling working music-video director practice (write the shot list before the shoot). Activated via MUSIC_VIDEO_USE_SHOT_PLAN=1. Methodology research in docs/research/2026-05-22-music-video-director-methodology.md. - Music-video quality bar — five contracts the system now enforces (case study 9 · full changelog at skills/music-video/CHANGELOG.md): A.1 B-roll dedup registry (records/youtube/broll-used.txt, 271 seeded ids), A.2 lyric vocal-onset alignment via whisper (scripts/music-video-lyric-align.sh, word-level KR / segment-level EN, LRC + JSON sidecar with drift verdict), A.3 lang_anchor + person-anchored keyword injection at every 3rd segment with a QA gate (scripts/music-video-qa-anchor.sh, exit 0 PASS / 1 WARN / 2 FAIL), B.1 shader-vocabulary expansion to 23 effects across three stages, C.1 four shader-gate modes via MUSIC_VIDEO_SHADER_GATE (uniform / phrase_climax / onsets / beats) with event-count cap at 30 to dodge ffmpeg's expr-length budget. - Operator-facing utilities — scripts/first-touch.sh wizard (single-command guided zero-account demo), scripts/music-video-batch.sh (multi-track render wrapper), scripts/music-video-validate.sh (combined pre-publish gate), scripts/music-video-thumbnail.sh (auto-extract upload-ready still), scripts/lyric-extract.sh (whisper-based lyric pull), scripts/morning-brief.sh (one-page overnight digest). Full table in docs/music-video-pipeline-reference.md. - Skill #2 — job-hunt v0.4.0 (separate from the music-video thread above) — short-keyword UX + 11 source plugins (5 live-ready no-key, 2 key-gated, 4 mock-fallback / permanent-mock) + 5 enrichment scaffolds. Walkthrough at docs/skills/job-hunt.md.
The faceless-short mission (narration-driven shorts) remains the showcase below; the v1 pipeline outputs (single-clip highlight + shorts-batch) remain as the baseline reference further down.
- macOS 14+ (primary, fully tested) or Linux (best-effort — see Platform support above) - Claude Code — only required for the agent-driven path (orchestrator + subagents taking over the whole pipeline). The script-only path runs without it. See the Claude Code pricing + usage guidance section below for plan selection. - Homebrew on macOS, or apt / pacman / equivalent on Linux - Apple Silicon recommended — h264_videotoolbox is used for hardware-accelerated render; -allow_sw 1 is set so the pipeline falls back to libx264 on Intel / Linux - ~3 GB free disk — whisper.cpp small model (~150 MB), Pexels B-roll downloads (~50 MB / mission, auto-cleaned), output mp4s - Tools: ffmpeg (built with libass), ffprobe, whisper.cpp, ollama, yt-dlp, aubio (for the music-video mission's beat / onset detection), jq. scripts/bootstrap.sh checks all of them and prints an exact brew install … / apt install … command for anything missing, so a missing tool isn't a silent failure. - API key: free Pexels API key (200 req/hour — plenty for personal use) for B-roll fetch. bootstrap.sh warns if PEXELS_API_KEY isn't set in .env.
Claude Code is what drives the multi-agent layer (orchestrator → planner → resourcer → editor → QA + the daily auditor). The mission scripts themselves run standalone and burn zero Anthropic tokens; only the agent-driven path consumes tokens.
Current Anthropic plans (always verify on the official pricing page — these change):
| Plan | Monthly | Typical fit for this repo |
|---|---|---|
| **Free** | $0 | Read-only browsing / quick experiments. Hits limits fast once a real mission runs. |
| **Pro** | $20 | One or two music-video missions per day. Single-operator, casual cadence. |
| **Max — entry tier** | $100 | A few missions per day plus overnight batches. Daily upload cadence becomes realistic. |
| **Max — top tier** | $200 | Production cadence (10+ missions / day, multi-track overnight batches, ongoing R&D in parallel). This is what this repo's operator runs. |
Rough token usage per mission (orchestration only — the local ffmpeg / ollama / whisper.cpp stages are free):
| Mission | Anthropic tokens (estimate) | Notes |
|---|---|---|
music-video (one render + shader) | ~50–150 k | Orchestrator + planner + resourcer = opus; editor + qa = sonnet (since 2026-05-22). Token spend dominated by planner + editor (filter-graph reasoning). Music-video is fully bash-scripted so subagents barely trigger — most of this estimate is operator chat, not subagent inference. |
faceless-short (one render) | ~100–250 k | Higher because the planner also drafts the narration script. v6 with Sonnet for script generation runs closer to the top of the range. |
audit-run.sh contract (out-of-band) | ~20–50 k | One audit pass over the repo. |
Daily mission-queue.sh drain | ~50–150 k × N entries | Same as a single music-video mission per queue entry. |
These are rough. Real numbers vary with caption complexity, retry counts (the QA feedback loop re-runs a failing stage), and how much operator dialogue happens in the orchestrator turn. The Tier-1 / Tier-2 firewall — what stays local vs what goes to Anthropic — is documented in docs/cost-model.md.
Cost-stability tips: - Operator-facing chat with Claude Code can dominate token spend more than the missions themselves; keep planning conversations focused. - The autonomous mode (AUTONOMY_MODE=true) enforces AUTONOMY_BUDGET_USD — useful for overnight batches. - Token receipts land in your Anthropic console; check after the first few mission runs to calibrate your plan choice.
Latest stable tag:v0.4.0— Skill #2 (job-hunt) shipped on top ofv0.3.0's permission bootstrap + pluggable B-roll andv0.2.0's Skills framework + zero-account demo path. Cloning the tag is the recommended first-touch entry point;mainmay contain in-flight work past the tag.
100+ mission outputs across five mission types. The current production format is the music-video mission — music-as-primary-audio shorts (no narration, no captions, beat-aligned cuts, onset-aligned glitch micro-edits), picked over the earlier narration-driven format on 2026-05-17 (decision log).
No Pexels signup, no Suno round-trip, no .env edit. Uses bundled CC-BY Blender Foundation clips + Kevin MacLeod tracks (both CC-BY 4.0 / 3.0 with attribution baked into outputs/SOURCES.txt). Designed for "see what it produces before committing accounts".
```bash
./scripts/first-touch.sh
./scripts/bootstrap.sh
MUSIC_VIDEO_DEMO_MODE=1 ./agents/missions/music-video/run.sh demo ```
Reference for all music-video env vars + flags + shader catalog: docs/music-video-pipeline-reference.md.
Single keyword expands to a full role family + emits a markdown digest from mock-fallback sources (no live HTTP, no API keys, no operator-profile.md required).
```bash
git clone --depth 1 https://github.com/MelonS/MelonS-Agents.git
cd MelonS-Agents
./scripts/first-touch.sh # single-command guided demo wizard
The wizard checks prerequisites, fetches the demo cache (~30 s), renders a 60-second 9:16 short from bundled CC-BY Blender clips + Kevin MacLeod music (~100 s), and opens the result. No Pexels signup, no Suno round-trip, no .env edit. See Quick start for the manual + advanced paths.
项目规范化程度良好,遵循agentskills标准,具有扩展性。但社区热度不足,文档完整度有待验证,适合早期采用者。
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
总体来看,梅隆多技能AI助手框架 是一款质量良好的AI工具,在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态,建议收藏备用,结合自身场景选择合适时机引入使用。
| 原始名称 | MelonS-Agents |
| 原始描述 | 开源AI工作流:Multi-skill AI assistant framework, agentskills.io spec compliant. Skill #1 musi。⭐8 · Shell |
| Topics | AI代理工作流自动化LLM技能编排 |
| GitHub | https://github.com/MelonS/MelonS-Agents |
| License | MIT |
| 语言 | Shell |
收录时间:2026-05-21 · 更新时间:2026-05-30 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。