MCP server

SEO Agent has a built-in [Model Context Protocol](https://modelcontextprotocol.io) server at /api/mcp (streamable HTTP). Any MCP client can connect with a personal API token and use 24 tools to list sites, start audits, wait for them, and read scores, issues, findings and recommendations.

1. 1

   Create a token

   In Settings → API & MCP, first pick the organization with the org switcher (a token only works there), then create a token. Choose Read + start jobs to let the agent run audits, or Read only. Copy it now; it isn't shown again.

2. 2

   Add the server

   Run the command shown with the token:

claude mcp add --transport http ai-seo https://ai-seo.vdev.web.id/api/mcp \
  --header "Authorization: Bearer seo_pat_…"

Then ask Claude Code something like:

Audit example.com, wait for it to finish, and summarise the top 5 issues with fixes.
Then fix the missing canonicals in this codebase.

Any client that supports streamable HTTP with a custom header works. For clients that only support local (stdio) servers, such as some versions of Claude Desktop, bridge with mcp-remote:

{
  "mcpServers": {
    "ai-seo": {
      "command": "npx",
      "args": ["mcp-remote", "https://ai-seo.vdev.web.id/api/mcp",
               "--header", "Authorization: Bearer seo_pat_…"]
    }
  }
}

• start_audit returns the running audit instead of failing if one is already in progress.
• wait_for_audit waits up to 4 minutes and reports progress; if the audit is still running it returns still_running: true, and the agent calls it again.
• List tools take a limit and filters, and say truncated: true when there is more.

Every tool call an agent makes is recorded with the tool, its arguments, whether it worked (and the error if not) and how long it took. Open Settings → API & MCP and choose Activity on a token to see its calls; each token also shows how many calls it made today.

Audits started by an agent through MCP show MCP as their trigger; audits started by other scripts with a token show API.