Evolink Router — Smart LLM Routing (Claude, GPT, Gemini, DeepSeek, Kimi)

Smart LLM routing brain for OpenClaw. Auto-dispatches tasks to Claude, GPT, Gemini, DeepSeek, Kimi via Evolink API. Cascade strategy cuts costs 60-85%. One A...

What to check before installing 1) Confirm Evolink trustworthiness: visit https://evolink.ai and review privacy/billing docs. The skill routes all requests through Evolink; that service will see prompts and could bill usage to the provided API key. If you proceed, create a scoped or separate API key if Evolink supports it. 2) Review and control config changes: the SKILL.md instructs merging changes into ~/.openclaw/openclaw.json (adding provider/models/agents). Back up that file first and manually inspect any JSON you paste in — do not blindly accept an automated merge. The registry metadata did not declare this config path. 3) Limit spawn privileges: the example config uses spawn:["*"]. Consider replacing that with a narrower list of models or denying spawn in your copy of the config if you want to limit what sub-agents can do. 4) Test with minimal scope: initially set the Evolink key with minimal billing limits or a test account and try read-only queries. Avoid sending secrets or sensitive data until you trust the routing behavior. The skill claims Evolink doesn't store requests long-term — verify this in Evolink's policy. 5) Be cautious about automatic file access: examples show the skill reading local files (CSV, project source). Only grant file access when asked for a specific task and verify which files are being read. 6) If unsure, ask for more info: request provenance (who published the skill), a link to the Evolink account/terms, and confirmation that the skill will not overwrite unrelated providers in your OpenClaw config. If the publisher cannot explain the config merge and spawn permissions clearly, treat the skill as higher risk.

Type: OpenClaw Skill Name: oc-skill-router Version: 2.0.0 The skill is classified as suspicious due to its broad permissions and inherent risks, despite lacking explicit malicious intent. The `SKILL.md` and `references/router-api-params.md` instruct the OpenClaw agent to configure itself with `spawn: ["*"]` permissions, allowing it to spawn any other registered agent. Furthermore, `references/cascade-examples.md` explicitly details tasks for spawned sub-agents that involve file system read/write operations (e.g., "读取 /data/sales-q2.csv", "在 /src/components/Sidebar.tsx 实现...") and command execution (e.g., "运行 TypeScript 编译检查"). While these capabilities are central to the skill's stated purpose of routing and task delegation, they create a significant attack surface for prompt injection, potentially leading to arbitrary file operations or remote code execution if user input is not adequately sanitized before being passed to spawned agents.