AI Visibility MCPPrivate beta
Run quoted AI visibility checks and read reports, baselines and fix plans from Claude, Cursor or any MCP client. These docs describe @s6s/visibility-mcp version 0.3.0.
What it is
Each tool is a call to the S6S AI Visibility API that returns the same confidence-labelled data you see in the Report, so the model you already work in can do the audit, the comparison and the proposal.
Two tools change something in your S6S account: register_brand creates a brand, and check_visibility runs a paid check. There are no tools that change a website: fix plans are suggestions your client approves and applies, and S6S never publishes anything.
Setup
1. Get a key
Access is by invitation. On s6s.ai, S6S issues the API key for your account (contact us); it can be limited to this MCP's own calls. readiness_snapshot works without a key; every other tool needs one.
2. Add it to your AI client
Add S6S to your MCP config (claude_desktop_config.json, .cursor/mcp.json, or your client’s equivalent). Set S6S_BASE_URL only for a self-hosted install; it defaults to https://s6s.ai.
{
"mcpServers": {
"s6s-ai-visibility": {
"command": "npx",
"args": ["-y", "@s6s/visibility-mcp"],
"env": { "S6S_API_KEY": "your-s6s-key" }
}
}
}The command starts the MCP server and waits for your client; call the tools from the client.
Version 0.3.0 is not on npm yet, so this command installs 0.2.0 for now. That version has no get_baseline_snapshot and can't confirm a new baseline: a check whose questions changed since the last run is refused and nothing is charged. Start that check with Run Check on the brand's page in S6S, which asks you to confirm.
Tools
readiness_snapshotFree · no keyA technical snapshot of any domain: whether the AI search crawlers OAI-SearchBot, PerplexityBot and Googlebot may read it, whether the homepage can be fetched and read without JavaScript, and which business details it states in code. It asks no AI engine and returns no score.
register_brandFree, or 10 creditsFinds the brand for a domain, or creates it, and always returns its brandId, which every other tool takes. An existing domain is free. A new domain runs one AI discovery step (name, category, keywords, competitors) for 10 credits. It does not start a check.
quote_visibility_checkFreeThe exact price of a check before anything is spent: credits required, your balance, the brand's daily limit, the questions, engines and repeats the check will use, and whether it would start a new baseline. It asks no engine and writes nothing.
check_visibility10 creditsRuns a full check and waits for it to finish. Pass the quoted credits as approvedCredits; if they don't match a fresh quote, or the balance, daily limit or a free-check rule doesn't allow it, it refuses and charges nothing. If the brand's questions changed, it refuses unless acknowledgeNewBaseline is true.
get_reportFree readThe latest Report: scores per kind of check with their confidence states, which engines named the brand, the answers behind them, and limitations.
get_baseline_snapshotFree readA short baseline from the latest check: how many answers named the business per AI company and in total, the date, the questions asked, which counts are early reads, and a share link to the full report that works for 30 days.
get_share_of_voiceFree readThe Share of Voice the Report shows: the brand's and each competitor's share of the business mentions counted in the answers (not a share of answers), the same split per question, other names the answers mentioned (unranked), and a quality label: insufficient, directional or reportable. Only a reportable result supports a rank claim.
get_fix_planFree readThe prioritized recommendations, each tied to the evidence behind it. They are suggestions the client approves and applies.
get_attributionFree readBefore and after for fixes already applied, with a confidence floor: only high-confidence movement is claimed; medium reads “likely contributed”.
One prompt is included: assemble_proposal (brandId, optional agencyName) pulls the report, share of voice and fix plan and writes a confidence-honest, client-ready proposal.
Quote, then approve
A check never spends credits nobody approved. The intended sequence is register_brand → quote_visibility_check → a person approves the quoted credits → check_visibility → get_report or get_baseline_snapshot.
- A check costs 10 credits; each brand can run up to 20 checks in any 24 hours.
check_visibilitytakes the approved figure asapprovedCreditsand refuses, with nothing charged, if it doesn't match a fresh quote.- If the brand's questions changed since its last check, the new results won't be comparable with earlier ones. The quote says so, and the check refuses unless you pass
acknowledgeNewBaseline: true. - Some partner accounts have free checks. Their quote shows 0 credits and the free checks left this month; each business gets one free full check.
- Registering a new domain is the one other step that can cost credits (10); ask before registering a domain S6S doesn't know yet.
The agency flow
Snapshot the prospect
Run readiness_snapshot on their domain: a free, instant read of whether AI search engines can read their site. Use it as the hook in the first conversation.
Register, quote, check
register_brand finds or creates them, quote_visibility_check shows the price, and once a person approves it check_visibility asks ChatGPT, Perplexity and Gemini the brand's buyer questions.
Show where they stand
get_report and get_baseline_snapshot show what AI said, kind of check by kind of check, with confidence labels intact. get_share_of_voice shows who AI names instead, with the Report's quality label.
Propose and follow up
get_fix_plan is the plan; assemble_proposal turns it into a client-ready proposal. After the client applies fixes, get_attribution shows what moved, within its confidence floor.
How to present the data
- Quote S6S’s scores and confidence labels exactly: an early read is never a confident number.
- Recommendations aim to improve readiness; they are not a guaranteed visibility lift.
- Every fix is applied by a person after approval. Nothing is published automatically.
- Only high-confidence movement may be claimed as attributed impact; medium is “likely contributed”.