AIMentionTracker MCP server

@aimentiontracker/mcp

Track how often ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews name and cite your brand, from inside your MCP client. Twelve read-only tools over AI visibility, share of voice, sentiment, competitor mentions and what to do about them, with the answers they came from.

This is answer engine optimisation (AEO) data, not rank tracking: the unit is an answer an assistant gave a real question, and whether you were in it.

Hosted (no install)

claude mcp add --transport http aimentiontracker https://mcp.aimentiontracker.ai \
  --header "Authorization: Bearer amt_live_YOUR_KEY"

Local (stdio)

claude mcp add aimentiontracker -e AMT_API_KEY=amt_live_... -- npx -y @aimentiontracker/mcp

Create a key at app.aimentiontracker.ai/api-keys. Keys are read-only and scoped to one organisation; they cannot change billing, the team, or what is tracked.

Tools

All read-only. v1 has no write endpoints, deliberately: adding a prompt expands a metered bill and triggering a run spends real vendor money.

Tool	What it answers
get_me	Who this key is, the plan, and the reporting thresholds. Call first.
get_current_date	Today in UTC, and the trailing windows every other tool counts back from. Call before reasoning about dates.
list_brands	Which brands this key can read, and their ids
list_questions	The questions this workspace runs, and how many answers each produced. Call before reading a disappointing number: it may be the wrong question, not poor visibility
get_visibility	Presence per engine and per brand over a window
get_share_of_voice	Each tracked brand's mentions as a share of the tracked set
get_sentiment	Recommended, negative and neutral counts per brand, against the classified denominator
list_competitors	Who gets named instead of you, including answers that named them and never mentioned you
list_answers	The stored answers, with mentions and citations
list_sources	Domains the engines read before answering
list_alerts	Findings worth acting on
list_recommendations	What to do next, ranked, with the pages and answers behind each item. Check displayable: false means the corpus is too thin to rank, not that there is nothing to do

One further tool, list_search_console, registers only where the Google Search Console integration is enabled. A tool that always answers 404 is worse than an absent one: an agent picks it for a task it cannot do and spends a turn failing.

Three rules the tools enforce

1. A null rate means "not enough answers yet", never zero. Rendering it as 0% reports a collapse that did not happen.
2. Engines are never averaged. They disagree; a mean describes no real surface.
3. An empty findings list does not mean nothing changed. list_alerts says so in the payload, not just the description.

Every tool that reports a measurement carries these rules in its own description, because an agent reads the tool, not the README.

Why the denominators are in the payload

get_share_of_voice returns brands_counted, and get_sentiment counts against classified rather than total mentions. A workspace tracking two competitors and one tracking twelve produce incomparable shares from identical answers, and the sentiment classifier abstains on listicle mentions where an engine named a brand without endorsing it. Both numbers are meaningless without the denominator beside them, so the tools return it rather than trusting the caller to ask.

Security

The hosted server holds no secrets. No database, no vendor credentials, no service account. Your key arrives in the Authorization header and is passed straight through to the API, which does all authentication, tenancy, entitlement and rate limiting. Each request gets a fresh, stateless server instance that is closed when the response ends.

More

• What it tracks and pricing
• Per engine: ChatGPT, Claude, Perplexity
• MCP server docs and the REST API

MIT.