How to Connect Promptwatch MCP to Claude, Cursor, and ChatGPT
We rank Promptwatch first for hosted streamable HTTP MCP in Claude, Cursor, and ChatGPT with OAuth or a Bearer key.
The ChatGPT plugin is not a different server than Cursor MCP. Plugin and Claude connector use the same MCP. Same project data. A visibility product you cannot ask about in Cursor stays a browser tab. The "same server" point is the one that decides the workflow, because a team that thinks the plugin is a separate integration builds two setups for one product, and two setups for one product is twice the maintenance for no extra data.
We rank Promptwatch first for MCP because the server is hosted streamable HTTP, the same URL works in Claude, Claude Code, Cursor, and ChatGPT/Codex, and those store listings are that server. Compare pipes in the directory. Review: Promptwatch. Product: promptwatch.com. Hosted streamable HTTP is the part that removes the local install, and the removed local install is the part that makes the connection a paste instead of a project.
MCP can read Overviews rows the project already stored. It does not explain Google's ranking of those answers. For that, use Google's AI features documentation. The chat is a client on Promptwatch data, the same way Slack is. The client analogy is the one that keeps the expectations right, because a client reads what the server has, and a server that has no prompts gives a client nothing to read.
Essential ($95/mo) includes MCP and API. Explore is free (10 ChatGPT prompts) and will not carry this. Professional is $245/mo. Business is $579/mo. Agency Kick-off $199, Growth $399, Scale $799. 4.7/5 on G2, 1,840+ brands.
Start Essential or above so MCP is included. We will not invent extra clicks, a local install, or a self-hosted binary. That is the whole connect path we will document: add the Promptwatch MCP URL to the client you already use. MCP does not crawl ChatGPT or the other engines itself. It reads what the project already stored. Without stored prompts and responses, the chat has nothing honest to say. A chat with no stored data is a chat that guesses, and a chat that guesses in a board deck is a liability.
Choose the key before you paste it
Auth is OAuth or a Bearer API key. Create a read-only project key and copy it once. A read-only key hides write tools, including generate content and CMS publish. Use this in a shared chat or a client workspace that must not publish. The read-only key is the part that turns a chat from a risk into a research tool, and the write tools are the part that the read-only flag removes.
Project key: one project. Org key: the org. If the room should only read, do not use a write-capable org key "for convenience." Add ChatGPT or the Claude connector only after the read path looks right. Use a write key later if you want actions from chat. The convenience argument is the one that causes the publish-no-one-meant, and the read-first path is the one that prevents it.
What the tools can do
Read (available even on read-only): visibility, sentiment, citations, Reddit and YouTube citations, prompts, competitors, gaps, fan-outs, site health, crawler, and visitors.
Write (only if the key allows): prompts, tags, topics, personas, generate content, CMS publish, reports, and actions.
The read list is the list a shared chat gets. The write list is the list a write key unlocks. The split is the part that lets a team share a chat without sharing the publish button, and the publish button is the part that should not be in a shared chat.
A first prompt we actually use: "Which prompts lost citations this week, and which pages did ChatGPTBot fetch but never cite?" That question needs the project data, not a pasted CSV. Ask for this week's citation losers and one crawler or visitor fact. If the project has no crawler pipe, the second half of that question will come back empty. That is a missing integration, not an MCP failure. The empty answer is the part that tells you the crawler pipe is missing, and the missing pipe is the part to fix in the app, not in the chat.
Paid monitoring is daily real UI on ChatGPT, Gemini, Claude, Perplexity, Grok, Llama, DeepSeek, Mistral, Copilot, AI Overviews, and AI Mode. MCP does not crawl those engines itself. It reads what the project already stored.
| Product | MCP we can rank | Notes |
|---|---|---|
| Promptwatch | Hosted HTTP; OAuth or Bearer; plugin and Claude connector | MCP on Essential |
| Otterly.AI | Mention tracking | $29; 4 engines; we will not invent an Otterly MCP |
| Peec AI | Tracking MCP | $95, 3 models; scores, not this write surface |
| Profound Starter | ChatGPT UI | $99/mo annual, ChatGPT-only |
Read the table by what is sourced. Promptwatch has a hosted HTTP MCP with OAuth or Bearer and the plugin and Claude connector are that server, on Essential. Otterly has mention tracking at $29 across four engines, and no MCP we will invent. Peec has a tracking MCP at $95 on three models, with scores and not the write surface. Profound Starter is $99 annual and ChatGPT-only. The sourced column is the one that decides what we will put on a slide, and an invented MCP is the one thing we will not put on a slide.
Peec's MCP is a real tracker pipe. We still rank Promptwatch first when the same chat needs citations, crawler logs, visitors, and (if allowed) CMS publish.
FAQ
Is the ChatGPT plugin a different server than Cursor MCP?
No. The plugin and the Claude connector use the same MCP. Same project data. Same hosted streamable HTTP URL.
Does Explore include MCP?
No. Essential ($95/mo) includes MCP and API. Explore is free (10 ChatGPT prompts) and will not carry this.
Can a read-only key publish to a CMS from chat?
No. A read-only key hides write tools, including generate content and CMS publish. Use that key in a shared chat or a client workspace that must not publish.