Key takeaways
- Promptwatch runs an MCP (Model Context Protocol) server that lets you ask ChatGPT, Claude, or Cursor questions about your brand's AI visibility data directly, no dashboard required.
- Setup takes about 10 minutes: generate an API key in Promptwatch, add the MCP server config to your client of choice, and start asking questions in plain language.
- MCP access is included on Promptwatch's paid plans alongside its REST API v2, and agencies can connect it per client for fast reporting.
- You can ask things like "What's my Visibility Score in ChatGPT this month compared to last month?" or "Which competitor is gaining share of voice in Perplexity?" and get an answer sourced from your real tracked data.
- This is different from Agent Chat (Promptwatch's own built-in analyst) in that MCP works inside tools you already use daily, like Claude Desktop or Cursor, rather than inside the Promptwatch dashboard.
I'll be honest, the first time I heard "MCP server for your AI visibility tool," my eyes glazed over a little. Another acronym, another setup doc. But once it's wired up, the appeal is obvious: instead of logging into a dashboard, filtering by date range, and exporting a CSV just to answer "are we losing ground to our competitor in Perplexity," you just ask. In Claude. Or ChatGPT. Right where you're already working.
This guide walks through what the Model Context Protocol actually does, how to connect it to Promptwatch, and what kinds of questions are worth asking once it's live.
What MCP actually is, briefly
Model Context Protocol is an open standard (originally from Anthropic) that lets AI assistants like Claude, or tools like Cursor, connect to external data sources and tools in a structured way. Instead of you copy-pasting numbers into a chat window, the assistant can query a live data source itself and reason over the response.
For a platform like Promptwatch, that means the assistant isn't guessing based on training data about "AI visibility trends." It's pulling your actual tracked prompts, visibility scores, citation data, and crawler logs, and answering based on that.
Why this matters for AI visibility specifically
AI visibility data is inherently conversational. You don't just want a number, you want to interrogate it. "Why did our score drop in August?" "Which prompts are we winning that we weren't three months ago?" "Is this a ChatGPT problem or a Gemini problem?" A static dashboard forces you to click through filters to answer each of these. An MCP connection lets you just ask, and the model can chain follow-up questions the way a human analyst would.
This fits into a bigger shift at Promptwatch toward agentic workflows. Its Agent Chat feature already does this inside the Promptwatch dashboard, but the MCP server extends the same querying capability to tools you already live in, like Claude Desktop, Cursor, or ChatGPT with developer mode enabled.
It's also worth remembering that Promptwatch already tracks AI coding agents like Claude Code and Codex as AI models in their own right, measuring whether they recommend your product when a developer asks them to plan an architecture. The MCP server is a separate, complementary thing: it's about querying your data through an agent, not about tracking agents as a visibility channel. Easy to conflate the two, worth keeping them separate in your head.
What you need before you start
- A Promptwatch account on a paid plan (Essential and above include API and MCP access; Explore, the free tier, does not)
- At least one project with tracked prompts and some response history, otherwise you'll be asking an empty database questions
- An MCP-compatible client: Claude Desktop, Cursor, ChatGPT (via its MCP/connector support), or any agent framework that speaks the protocol
If you haven't set up tracking yet, do that first. Create a monitor for your core category, add 20 to 50 prompts that mirror how real buyers phrase questions, and let it run for at least a week before you start querying it through MCP. Otherwise you're asking Claude to analyze a trend line with one data point.
Step-by-step setup
Step 1: Generate your Promptwatch API credentials
Log into your Promptwatch dashboard and head to your account or workspace settings. Look for the API and integrations section, where you'll find the option to generate an API key tied to your account or a specific project. Treat this key like a password. It grants read access to your tracked visibility data, so don't paste it into a public repo or a shared Slack channel.
If you're an agency managing multiple clients, you can generate separate keys per client workspace, which keeps the data cleanly scoped when you're querying through an assistant later.
Step 2: Locate the MCP server endpoint
Promptwatch exposes its MCP server alongside its REST API v2. The exact connection details (server URL and authentication method) are documented in your dashboard's developer section. Copy the server URL, you'll need it for the client-side config in the next step.
Step 3: Add the server to Claude Desktop
If you're using Claude Desktop, open its settings and find the Developer or MCP Servers section. You'll add a new server entry with:
{
"mcpServers": {
"promptwatch": {
"command": "npx",
"args": ["-y", "@promptwatch/mcp-server"],
"env": {
"PROMPTWATCH_API_KEY": "your-api-key-here"
}
}
}
}
Save the config and restart Claude Desktop. You should see Promptwatch listed as an active connector, usually with a small plug or tool icon in the chat interface confirming it's live.
Step 4: Add the server to Cursor or ChatGPT
Cursor supports MCP servers through its own settings panel, where the process is nearly identical: paste the server command, add your API key as an environment variable, and restart. ChatGPT's support for external MCP connectors is rolling out through its developer and connector settings, check your account type, since availability varies by plan.
Step 5: Run a test query
Once connected, ask something simple first, like "List the projects I have in Promptwatch." If that comes back correctly, you know the connection is live and authenticated. Then move on to real questions.
What to actually ask it
This is the part people skip past, and it's the most important part. An MCP connection is only as useful as the questions you bring to it. Here are the categories worth building a habit around:
- Visibility trend questions: "What's our Visibility Score across ChatGPT and Gemini for the last 30 days, and is it trending up or down?"
- Competitive questions: "Which competitor gained the most share of voice in Perplexity this month, and on which prompts?"
- Diagnostic questions: "Pull the prompts where we dropped out of the top three mentions in the last two weeks."
- Crawler and technical questions: "Has ClaudeBot hit any errors crawling our pricing page in the last 7 days?"
- Content gap questions: "What topics is our top competitor getting cited for that we're not covering?"
Notice none of these are yes-or-no questions. The value of routing this through an assistant instead of a dashboard filter is that you can ask a follow-up in the same breath: "okay, now break that down by model" or "show me the actual AI response text for three of those."
MCP vs. Agent Chat vs. the dashboard: when to use which
People sometimes assume MCP replaces the Promptwatch dashboard entirely. It doesn't, and it shouldn't. Each interface is built for a different moment.
| Interface | Best for | Where it lives |
|---|---|---|
| Promptwatch dashboard | Visual trend review, setting up monitors, approving Content Agent drafts, scheduled reports | promptwatch.com |
| Agent Chat | Deep multi-step analysis with charts, inside Promptwatch, including Slack | Promptwatch app and Slack |
| MCP server | Quick ad hoc questions from inside Claude, Cursor, or ChatGPT, without switching tabs | Your existing AI tools |
| REST API v2 | Custom dashboards, Looker Studio, automated pipelines | Your own codebase or BI tool |
If you're doing deep strategic planning, say, building a Q4 content roadmap, the dashboard with its Answer Gap Report and Unified Actions list is still the right place. If you're mid-conversation in Claude and want a quick number to back up a Slack message, MCP is faster. They're complementary, not competing.
Common setup mistakes
The API key gets pasted with a trailing space or quote mark more often than you'd think, and the connector will fail silently or throw a vague auth error. Double check it pasted cleanly.
Another one: people connect MCP before they've actually set up prompt tracking in Promptwatch, then wonder why every query comes back empty. The MCP server surfaces your data, it doesn't generate data on its own. Get your monitors running first.
And a smaller one, but it trips up agencies specifically: if you've got multiple client workspaces under one Promptwatch account, make sure you're scoping your API key (and therefore your MCP queries) to the right project. Nobody wants to accidentally report a competitor's visibility numbers as their client's.
For agencies: a faster reporting loop
If you're running AI visibility reporting for multiple clients, the MCP server changes the shape of a weekly check-in. Instead of exporting a PDF and talking through it on a call, you can open Claude, connect the right client's Promptwatch key, and ask live questions during the meeting. "Pull up their ChatGPT visibility trend since we published that comparison page" gets you an answer on the spot, not a promise to follow up by email.
This is the same philosophy behind Promptwatch's whitelabel dashboards and client portal, just extended into a conversational layer. Agencies managing Monks-style enterprise accounts or smaller retainer clients both benefit from cutting the lag between "client asks a question" and "client gets an answer."
A word on data accuracy
One thing worth knowing before you lean on this day to day: Promptwatch pulls its visibility data from the real interfaces of ChatGPT, Gemini, Perplexity, and the rest, not just their APIs. That matters here because API-only tools can return different answers than what a real user sees in the actual chat app. When you ask Claude a question through the MCP connection, you're getting numbers grounded in what buyers actually saw on their screens, not a synthetic API response. Promptwatch's own research on citation volatility, like its tracking of the ChatGPT citation drop after the GPT-5.3 rollout (https://promptwatch.com/data/chatgpt-citation-drop), is a good example of why that distinction matters. Numbers move fast in this space, and a tool reading from the wrong source can mislead you about whether a dip is real or just an artifact of how it's measuring.
If you're evaluating other AI visibility tools first
MCP access is a nice detail, but it's not the reason to pick a platform. If you're still comparing options, look at the fundamentals: how many AI models it tracks, whether it has crawler log access, whether it generates content or just reports problems. The directory at bestgeosoftware.com is a reasonable place to see how different GEO platforms stack up on those fundamentals before you commit to one and wire it into your workflow.
Once you've picked a platform, though, the MCP layer is worth setting up early. It costs ten minutes and changes how often you actually check your numbers, because checking becomes as easy as typing a sentence instead of opening a new tab.
Frequently asked questions
Do I need a developer to set up the Promptwatch MCP server?
No. The config is a short block of JSON pasted into a settings panel. If you can edit a config file and paste an API key, you can do this yourself in under 15 minutes.
Which Promptwatch plans include MCP access?
MCP and API access are included starting on the Essential plan ($95/month), alongside Professional, Business, and the agency tiers. The free Explore plan doesn't include it.
Can I connect MCP to more than one AI assistant at once?
Yes. You can add the same Promptwatch MCP server config to Claude Desktop, Cursor, and other MCP-compatible clients simultaneously, each with its own API key if you want separate access scopes.
Does MCP access replace the need to look at the Promptwatch dashboard?
No. MCP is for fast, conversational questions. The dashboard is still where you set up monitors, review Content Agent drafts, approve Unified Actions, and look at visual trend charts. Think of MCP as a quick-reference layer on top of the same underlying data.