Install OpenClaw skills and MCP servers for SEO safely

The registry is open, so you are the reviewer

ClawHub, OpenClaw’s public registry for skills and plugins, is open by default: anyone can upload, and the documentation does not claim that its automated scans guarantee safety (ClawHub docs). The Hacker News reported, citing Koi Security, 341 malicious skills among 2,857 audited (the report). This guide turns that into a vetting workflow for SEO work. We name no skill or plugin, because we have not reviewed any.

A vetting workflow, step by step

1. Read before you enable. The documentation says to treat third-party skills as untrusted code, read them first and keep credentials out of skill content (skills docs).
2. Check the registry page. Public detail pages summarise the latest scan state, and signed-in users can report a skill. A clean summary is not proof of safety.
3. Install from a reviewed copy. There is no documented version-pin syntax for registry installs, but you can install a local directory with openclaw skills install ./path --as name. Review the folder, then install that copy.
4. Check for name shadowing. When two skills share a name, the one from the higher-precedence location wins, and workspace skills rank above bundled ones. A same-named skill in the workspace silently replaces a trusted one.
5. Keep secrets out of the skill. Per-skill env and apiKey are applied to the host process for a run and then restored; they do not reach a sandbox.
6. Turn off optional telemetry. When you are logged in, the clawhub CLI may send a best-effort install event; CLAWHUB_DISABLE_TELEMETRY=1 disables it.
7. Record what you installed. The CLI writes versions to .clawhub/lock.json; commit that file.

Adding an MCP server

Saved servers live under mcp.servers.<name>. A remote server takes url, auth and timeouts; if you omit transport, OpenClaw uses sse, and "transport": "streamable-http" selects streamable HTTP (transports docs). That matters: Semrush documents that its server uses streamable HTTP only (Semrush MCP docs), so set the transport explicitly for it and for any server documented the same way.

{
  "mcp": {
    "servers": {
      "search-console": {
        "url": "https://mcp.example.com/mcp",
        "transport": "streamable-http",
        "auth": "oauth",
        "requestTimeoutMs": 30000,
        "toolFilter": { "include": ["read_*"] }
      }
    }
  }
}

An example shape only, with a placeholder address. With auth: "oauth" a static Authorization header is ignored; sign in with openclaw mcp login <name>. Use openclaw mcp status to inspect saved config without starting servers, and probe to open a live connection.

Give the SEO agent the smallest tool set

Tools from connected servers go through the same tool profile and policy as built-in ones, and a per-server toolFilter.include or exclude trims what a server exposes (MCP docs). For a reporting agent, include only read tools, as in the example. You can also deny individual tools for one session. Put credentials in the supported secret mechanisms; openclaw mcp doctor warns when headers or env hold literal sensitive-looking values.

A minimal SEO setup

OpenClaw skills and MCP for SEO, answered

Are ClawHub skills reviewed before they go live?

The registry runs automated checks, but it is open by default and anyone can upload; review what you install.

Can I pin a skill version?

Installs record versions in a lock file, but the documentation shows no pin syntax; install a reviewed local copy instead.

Why does my streamable HTTP server not connect?

If the transport is omitted OpenClaw uses sse; set "transport": "streamable-http".

What to read next

Read ten rules for running an agent on Search Console data and continuous analysis with an AI agent.