Connect your AI assistant to sirefs

sirefs shows what to do about your AI visibility: the prompts where competitors are named and you aren't, the pages to improve, the sites to get cited by, the audit checks to fix. An assistant that can read those steps can also do the work: draft the page, write the outreach, compare this week with last. The MCP server is how it reads them.

What is MCP?

The Model Context Protocol is the open standard assistants use to call tools in other products. Claude, ChatGPT, Cursor and most coding agents support it. A server lists its tools; the assistant picks one when a question needs it, calls it, and reads the result. sirefs' server is at https://api.sirefs.com/mcp and takes a workspace API key.

What can the assistant do?

Read, on any project in the workspace:

• the overview: visibility, citation rate, recommendation rate, rank, sentiment and share of voice for your brand and each competitor, in total and per engine, every rate with its sample size and interval;
• the prompts with the brand's visibility per engine, and the trend over time;
• competitors head to head, and the untracked brands the engines name most;
• the pages and domains the engines cite, and the community threads among them;
• the GEO audit with every check's finding and fix, at your plan's level;
• the next steps, ranked by impact, with their evidence;
• recent alerts, and any collected answer in full.

Change, only when you ask it to: add or pause prompts, start tracking a competitor, and run the audit again. The same plan limits apply as in the app: a prompt over your allowance is refused with the reason, and the assistant tells you. The server never starts anything your plan doesn't already cover.

How do I connect?

Give your assistant the server address, https://api.sirefs.com/mcp. It sends you to sirefs to sign in and choose the workspace, and it's connected. There is no key to copy.

Claude (claude.ai and Claude Desktop). Settings, Connectors, Add custom connector. Paste the address and continue; sign in when asked.

ChatGPT. Settings, Connectors, Create. Paste the address, choose OAuth as the authentication, and sign in when asked.

Claude Code.

claude mcp add --transport http sirefs https://api.sirefs.com/mcp

Then type /mcp in Claude Code and choose sirefs to sign in.

Cursor and other editors. Add the server to the editor's MCP settings; it opens the sign-in itself:

{
  "mcpServers": {
    "sirefs": { "url": "https://api.sirefs.com/mcp" }
  }
}

The sign-in is standard OAuth 2.1 with PKCE; the assistant registers itself as a client, so one that isn't listed here works too. Connected assistants are listed under Settings, API in the app, where anyone in the workspace can disconnect their own and owners and admins can disconnect any.

Scripts and tools that can't sign in take an API key instead, made under Settings, API, and sent as a bearer token: Authorization: Bearer sir_.... A key never expires and stands for the whole workspace. The same key works for the REST API.

What should I ask?

• "What should we work on this week for Acme?" The assistant calls get_next_steps and explains each step with its evidence.
• "Which prompts name our competitors but never us, and which pages do we have for them?" get_prompts, then get_next_steps for the pages the audit found.
• "Draft the page you'd write to get cited for the top missing prompt." It reads the answers the engines gave (list_answers, get_answer) and the sources they cited, and writes from those.
• "Did Perplexity change this month?" get_trend per engine, with sample sizes.
• "Start tracking the brand that keeps showing up." get_competitors for the suggested list, then track_competitor once you confirm. If it's a source you rely on rather than a rival, mark_not_competitor keeps it out of the suggestions.

Rates on small samples are rough; the assistant is told the sample size with every rate and should say so. If it gives a number without one, ask for it.

Access and limits

• A connection belongs to the person who signed in and the workspace they chose; the assistant sees every project in that workspace and changes them as that person could. Access tokens last an hour and renew themselves; disconnecting ends them at once.
• A key stands for the workspace. Whoever has it sees every project in it, so keep it where you keep passwords, and make one key per tool so you can revoke one without the others. Owners and admins make and revoke keys.
• The MCP server is included in Pro and Agency. A connection made on Pro keeps answering after a downgrade, but every call then explains that the plan doesn't include it.
• Each connection or key may make 120 calls a minute, more than a conversation needs.
• The tools read what's already collected. They never ask an AI engine on your behalf, so a chat with your assistant costs nothing against your plan beyond the prompts you add.

Which tools are there?