kie

claude-plugin-kie

A Claude Code plugin that installs a kie.ai concierge agent. When the conversation naturally arrives at "let's generate this with kie.ai", Claude delegates to this specialist — it recommends models, checks credits, builds the request from bundled docs, calls the API, and returns the result.

kie.ai is a unified API platform for 100+ AI models: Suno (music), Runway/Veo3/Kling/Wan/Sora2 (video), Flux/Ideogram/Seedream (image), ElevenLabs (audio), Claude/Gemini/GPT-5 (chat), and more.

How it works

This plugin is not a slash command. It installs a specialist subagent (kie-agent) that Claude delegates to when appropriate:

User: "この動画にBGMほしいな"
Claude: "kie.aiで作れますね、候補を確認します"
  └─ (internally: Claude → kie-agent "suggest models for BGM")
Claude: "Suno V5 と ElevenLabs sound-effect があります…"

User: "Sunoで、2分、ジャズ系、歌詞なし"
Claude: "残クレジット確認しますね"
  └─ (internally: Claude → kie-agent "check credits")
Claude: "5,243 credits 残っています。実行していい？"

User: "うん"
  └─ (internally: Claude → kie-agent "execute Suno generate-music with ...")
Claude: "生成完了: https://..."

What's inside

claude-plugin-kie/
├── .claude-plugin/plugin.json  # plugin manifest
├── agents/kie-agent.md          # the concierge subagent
├── scripts/                     # shared runtime
│   ├── call.sh                  # POST /api/v1/jobs/createTask
│   ├── poll.sh                  # poll recordInfo until done
│   ├── upload.sh                # file upload helper
│   ├── query_graph.sh           # query knowledge graph (wrapper)
│   └── query_graph.py           # NetworkX-based BFS/DFS over graph.json
├── docs/                        # snapshot of docs.kie.ai (194 markdown files)
├── graph/                       # pre-built knowledge graph via graphify
│   ├── graph.json               # 1,107 nodes / 1,341 edges
│   └── GRAPH_REPORT.md
└── reports/                     # architecture analysis

Why bundle docs + graph?

kie.ai adds new models frequently. The bundled docs/ is a pinned snapshot, and graph/graph.json is a pre-computed knowledge graph (via graphify) that lets the agent find the right endpoint ~69x more token-efficiently than reading raw docs.

See reports/api-response-schema-analysis.md for why kie.ai's unified ApiResponse schema enables shared runtime scripts across all 100+ models.

Setup

1. Install prerequisites

• uv — required for the graph query helper to auto-install its Python dependency (graphifyy/networkx). See uv installation guide.
• jq and curl — used by the runtime scripts for API calls and JSON parsing.

On first use, scripts/query_graph.sh will auto-run uv tool install graphifyy if graphifyy is missing. No manual step required.

2. Install the plugin in Claude Code

/plugin marketplace add https://github.com/hirakawat-hmp/claude-plugin-kie
/plugin install kie@claude-plugin-kie
/reload-plugins

3. Store your kie.ai API key outside the plugin

mkdir -p ~/.config/kie
echo 'export KIE_API_KEY=your_key_here' > ~/.config/kie/env
chmod 600 ~/.config/kie/env

The agent's scripts source this file — the API key never enters the Claude context and is never committed to the repo.

Refreshing the documentation

kie.ai adds new models frequently. To refresh the bundled docs and rebuild the knowledge graph, the agent includes a refresh-docs procedure (TBD — likely triggered by explicit user request like "kie.aiのドキュメント更新して").

Status

Work in progress. Current state:

• ✅ Docs snapshot (194 files) bundled
• ✅ Knowledge graph built
• ⏳ agents/kie-agent.md (implementation)
• ⏳ scripts/call.sh and scripts/poll.sh
• ⏳ End-to-end testing

License

MIT