Claude Code Pack

Copy-paste only — no code to write. Once installed, tell Claude Code: "看下今天客服有什么要处理的" ("show me today's pending support work").

Install steps

1. Create a Gateway Key in the KefuAgents settings panel (ecosystem: Claude Code; default scopes gateway:read + gateway:submit_draft). The key is shown once — store it in the KEFUAGENT_GATEWAY_KEY environment variable, never in a committed file.
2. Save the mcp.example.json content below as .mcp.json in your project root (or register with the one-line CLI alternative).
3. Install SKILL.md into your Claude Code skills directory (user-level ~/.claude/skills/kefuagent-support/SKILL.md or project-level .claude/skills/kefuagent-support/SKILL.md — confirm against your Claude Code release; pending validation).
4. Ask Claude Code: "看下今天客服有什么要处理的".

.mcp.json (raw file: /agent-packs/claude-code/mcp.example.json)

{
  "mcpServers": {
    "kefuagent": {
      "type": "http",
      "url": "https://kefuagents.com/api/gateway/mcp",
      "headers": {
        "Authorization": "Bearer ${KEFUAGENT_GATEWAY_KEY}"
      }
    }
  }
}

• The ${KEFUAGENT_GATEWAY_KEY} env-var expansion keeps the key out of committed files — the exact expansion syntax is pending validation against the current Claude Code release; if unsupported, use the CLI form below and never commit a .mcp.json containing a literal key.

CLI alternative (flag names pending validation):

claude mcp add --transport http kefuagent https://kefuagents.com/api/gateway/mcp \
  --header "Authorization: Bearer $KEFUAGENT_GATEWAY_KEY"

SKILL.md (raw file: /agent-packs/claude-code/SKILL.md)

The skill is maintained in Chinese — the merchant supervision language. It teaches the five-action supervised workflow: list cards → fetch reply context → draft with the session's own model → submit behind the safety gate → escalate high-risk items to humans. Download it from the raw link above; the canonical content is embedded on the Chinese page.

What the skill teaches (and the platform enforces anyway)

The skill teaches the five-action supervision model: read cards, read context, submit drafts, direct rewrites in Chinese, escalate to humans. Even if an agent ignores every instruction, the platform side still holds: the gate re-checks every draft, high-risk stays human, one key = one workspace, every call is audited.

Plugin-marketplace publication is a follow-up distribution step; this page documents manual installation.