Check your own AI visibility from Claude or Cursor: the MCP workflow

You can run an AI-visibility scan on any domain from inside Claude Code, Claude Desktop, Cursor, Windsurf, or VS Code. Collimer ships an MCP server, collimer-mcp v0.3.0 as of August 2026, that exposes exactly one tool, collimer_scan, installed with a five-line JSON config block and run on demand via npx with no separate install step.

The tool returns a score from 0 to 100 with a confidence interval, the single biggest visibility gap, a link to the full scorecard, and an unlock link. It does not return the full ranked fix plan or the verification re-scan; those require a free account on the web. That boundary is intentional, and it is explained in the section on where the agent stops.

What can the Collimer MCP server actually do?

The server exposes exactly one tool: collimer_scan(domain, email?). Pass a domain; the email parameter is optional and links the result to a free account so you can claim the full fix plan.

What comes back in the agent’s response:

• A score between 0 and 100 with its confidence interval, so the number your agent returns is the same measured number, with the same honesty about noise, you would see in the browser.
• The single biggest AI-visibility gap the scan found, in plain language.
• A link to the full scorecard, and an unlock link to claim the ranked fix plan with a free account.

What does not come back:

• The ranked list of all gaps and fixes.
• A verification re-scan confirming a fix moved the score.
• Any write access. The agent reads and reports. It does not edit your pages, submit your sitemap, or push changes anywhere.

The tool calls the same probe engine as the web free scan, described in how to measure AI visibility properly. The score is the same scan, surfaced closer to where you work.

How do I install it in Claude Code, Claude Desktop, Cursor, or VS Code?

Add this block to the right config file for your client. The block is identical across clients; only the file path changes.

{
  "mcpServers": {
    "collimer": {
      "command": "npx",
      "args": ["-y", "collimer-mcp"]
    }
  }
}

File paths by client, as of August 2026:

• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows
• Claude Code: .mcp.json in the project root
• Cursor: ~/.cursor/mcp.json
• Windsurf / VS Code: the editor’s MCP settings, or .vscode/mcp.json in the project root

No global install is needed: npx -y collimer-mcp pulls the package on first run and caches it. To pin the version, use "collimer-mcp@0.3.0".

Five optional environment variables tune behavior if you hit timeouts on slow connections: COLLIMER_API_BASE (default https://app.collimer.com), COLLIMER_SCAN_SOURCE (attribution tag, default mcp), COLLIMER_MAX_POLLS (default 60), COLLIMER_POLL_INTERVAL_MS (default 5000), and COLLIMER_REQUEST_TIMEOUT_MS (default 30000). For most setups, none are needed: add the block, restart your client, and the tool is available.

What should I ask it?

Three prompts from the collimer-mcp README (v0.3.0, August 2026) work verbatim:

1. “How visible is stripe.com in AI search?”
2. “Run a Collimer scan on example.com and tell me the biggest AI-visibility gap.”
3. “Does ChatGPT cite acme.com when asked about its category?”

For your own domain, the most useful follow-up after the scan is to ask the agent to explain the gap in plain terms and suggest what kind of content change would address it. The agent can reason about the gap using what it knows about how five AI engines decide what to cite. It cannot apply the fix, but it can help you draft the paragraph or the structured-data block that would.

What does a real session look like?

This is an unedited transcript from a session on August 24, 2026, scanning our own domain. We are publishing the real number rather than a flattering one, because that is the whole point of measuring.

Prompt entered:

Run a Collimer scan on collimer.com and tell me the biggest AI-visibility gap.

Tool call made by the agent:

collimer_scan(domain="collimer.com")

Returned by the tool, verbatim:

AI-search visibility for Collimer: 3/100 (±4).
Biggest gap: Collimer is absent from category-awareness prompts (0/16), meaning
LLMs do not associate it with the named category 'AI search visibility & brand
citation platform', creating an entity-definition gap that prevents it from
appearing when buyers ask what tools exist in this space.
Full scorecard: https://app.collimer.com/scan/[redacted]
Unlock the full ranked fix plan for this domain with a free account:
https://app.collimer.com/users/register?scan=[redacted]
Tip: after you make changes, re-run the scan to measure the delta.

That is the complete output: a score with its interval, one gap in plain language, two links, and a nudge to re-measure after shipping. The tool did not invent extra findings, did not return a score without its interval, and did not pretend it could apply the fix. And yes, 3 out of 100 is our own score on category prompts as of that scan — August 24, 2026. Treat it as a dated baseline, not a verdict: this is day zero of a public experiment. We are working the fix plan it points at in the open, re-probing the same frozen prompt panel at each checkpoint, and publishing the deltas with their intervals — including the fixes that do not move anything. If you quote the 3, quote the date.

Where does the agent stop and the human start?

The agent returns one gap. The full ranked fix plan lives behind a free account at app.collimer.com; for our own domain’s reports it has run to a dozen prioritized items per scan. Creating the account takes under two minutes and links the scan you already ran to your profile.

The verification re-scan, which confirms a fix actually moved the score, is also behind the account. Verification is metered because each re-check runs real probes against live AI engines; the report shows what a re-check involves before you run one.

If you want to call the scan programmatically without MCP, the OpenAPI spec is at app.collimer.com/api/v1/openapi.json and the interactive docs are at app.collimer.com/docs. The same scan is available as a REST endpoint for teams building their own tooling or CI checks.

What are the limits?

Three limits are worth knowing before you hit them, all as of August 2026:

• Per-IP daily cap. The free scan allows two scans per IP address per day. A third scan from the same address is refused immediately, not queued. On a shared office network or a VPN exit node, other people’s scans count against the same cap.
• Per-email and global caps. A separate cap applies when you pass the email parameter, and a global daily budget applies across all anonymous scans. On high-traffic days the global budget can run out before your per-IP cap does.
• Poll timeouts. The scan is asynchronous; the tool polls until the result is ready or its poll budget is exhausted. If it times out on a slow connection, the scorecard link is still valid and the full result lands there within a few minutes.

What the server does not do: it does not write content, push commits, update structured data, or submit anything to any engine. The workflow is read-only on the agent side. Every fix is a human action.

One thing to do today

Add the five-line block to your .mcp.json, ask your agent “How visible is <yourdomain> in AI search?”, and paste the returned biggest gap into your next standup. Then open the scorecard and claim the ranked fix plan: the scan from your editor is the same scan, just closer to where you work.