The short answer
Paste the setup prompt into an AI app that supports remote MCP servers. Your agent gets DuckDive connected (a terminal agent registers the server itself; a chat app hands you its add-connector link), waits while you sign in, then checks your setup status, adds your website, links whichever GA4 property and Search Console site match your domain without doubt and hands you a dashboard link. It never sees or types your password.
- Your one sign-in step: Continue with Google, or Allow if you are already signed in.
- The agent asks before moving a data source that another website already uses.
- Only DuckDive admins can set up. A team member's agent reads data and points you to your admin.
What happens after you paste the prompt
The prompt points your agent at agents.md, a plain-text guide with a section written for the agent: the server address, the order of calls, what to ask and what never to do. The top half is for people (the prompt to paste, connect steps per app, and a route without AI). Your agent reads the whole file, then works through setup in one sitting and narrates as it goes.
What it is setting up: DuckDive keeps a separate store for each website, filled from four Google products (GA4, Search Console, Ads and Local Services Ads) plus Yext and Local Falcon, and your agent reads that store over MCP. Setup is simply telling DuckDive which site is yours and which Google properties belong to it.
The run, in the guide's order:
- If DuckDive tools such as
get_setup_statusare missing, it registershttps://api.duckdive.ai/mcpitself where it can (Claude Code, Gemini CLI), or gives you the add-connector link or steps for your app, with no API key either way. - A DuckDive page opens in your browser and the agent asks you to select Continue with Google. That is your part.
- It calls
get_setup_statusand readsnext_step.code, which says where your account stands. - It adds your site with
add_website, using the address you named or the websites DuckDive found in your Google account. - It calls
list_available_sources, links every clear match withconnect_website_sources, and tells you what it linked. - It gives you your dashboard link, explains that the first import is normally a matter of minutes, though years of history stretch it, and asks what you want to know.
The guide tells the agent to ask only when the answer is ambiguous, and then once, as a short numbered list. In practice you hear from it when two suggested websites or two Search Console properties could both be right, and it gets on with the work otherwise.
Name your site and skip a question
If you already know which address to add, put it in the sentence. The agent then has nothing to choose between and goes straight to linking.
Paste into your agentSet up DuckDive for my website, example.com. Follow https://analytics.duckdive.ai/agents.md
Tools it calls get_setup_status add_website list_available_sources connect_website_sources get_sync_status
Swap example.com for your domain and leave the link alone; it is where the agent's instructions live.
Check these before you paste it
- An app that takes remote MCP servers. Claude, Cursor, Claude Code and VS Code are documented; DuckDive says ChatGPT needs Business or Enterprise, and Gemini custom apps need a personal Google account in the US.
- The right Google login in your browser. Whichever account you click with decides which GA4 properties and Search Console sites DuckDive can offer.
- Half an hour free. The DuckDive sign-in page expires after 30 minutes; if it lapses, reconnect from your AI app.
- An admin seat, if your team already uses DuckDive. Agency Admin or Client Admin can set up. Team members can only read.
- The client name, if you are an agency. Multi-client accounts must place each site under a client; the agent asks if you leave it out.
The one click that stays yours
No agent finishes this setup without you, and that is deliberate. Signing in to Google happens on Google's own screen in your browser, and the guide forbids the agent from asking for, storing or typing passwords or Google credentials. In a chat app such as Claude on the web or ChatGPT you also press the app's own Add or Connect button first, because a chat agent cannot install its own connector; the sign-in is the step no app lets you skip.
When the agent connects, a DuckDive page opens. You see one of two screens:
- New to DuckDive: select Continue with Google. One Google screen signs you in, creates the account with no form to fill, shares your Analytics, Search Console and Ads data, and approves the agent's connection.
- Already signed in to DuckDive: an approval screen appears. Leave the Also allow actions box checked if the agent should be able to set things up, then select Allow.
Why this step cannot be handed off: remote MCP servers like DuckDive's sign clients in with OAuth. In the flow drawn in the protocol's authorization chapter, which we read on November 16, 2026, the client opens your browser at an authorization URL, you approve there, and only then does the authorization server issue a token; the chapter bases this on an OAuth 2.1 draft. Your AI app ends up holding a DuckDive access token, never your password.
If the DuckDive tools do not show up after the click, turn DuckDive on in that chat's tools menu or open a new chat. A sign-in that keeps looping usually clears once you log in to DuckDive in that browser before connecting.
Each setup state your agent moves through
get_setup_status is the agent's map. It returns your role, whether this connection may make changes, your clients, your websites with each source's import status, whether Google is connected, websites suggested from your Google account, a next_step with a code and a message, and a set of links. The agent acts on the code. There are six:
add_website- No website yet. The agent adds the one you named or the lone suggestion, offers a numbered list when there are several, or asks for your address.
connect_google- Google is not connected, usually after an email sign-up or a cancelled Google screen. The agent sends the connect link and waits for "done".
choose_sources- A website exists without linked data. The agent lists what your Google logins can reach and links the clear matches.
importing- Sources are linked and history is coming in. The agent gives you the dashboard link and can track progress with
get_sync_status. ready- Data is in. Setup is over and questions start.
ask_admin- Your seat is a team member's, which cannot set things up. The agent stops and tells you to ask your DuckDive admin.
You can pull up the same map any time: it is the fastest way to see why a site shows nothing, or to resume a half-finished setup.
Ask your agentCheck my DuckDive setup status and tell me what step is next.
Tools it calls get_setup_status get_sync_status
Worked example: a setup run with two suggested websites
Lanternfish Studio is a made-up two-person design studio with a portfolio site and a separate print shop. The owner clicks in with the Google account that holds both GA4 properties and pastes the setup prompt without naming a site. Years ago the studio verified its Search Console twice, once as a domain property and once as a URL-prefix property, which is the wrinkle this run has to deal with.
Set up DuckDive for my website. Follow https://analytics.duckdive.ai/
get_setup_statusadmin seat, actions allowed, Google connected, no websites yet, two suggested websitesadd_websiteadds lanternfish-studio.example once the owner picks itlist_available_sourcesone unused GA4 property in ready_to_link; two Search Console properties fit the domain, so that source is marked unclearconnect_website_sourceslinks the GA4 property and starts its history importconnect_website_sourceslinks the Search Console property the owner choseget_setup_statusthe next morning, when the owner asks what step is next
"Your Google account shows two websites. Which should I add? 1. lanternfish-studio.example 2. lanternfish-prints.example" (The owner typed 1.)
"Added lanternfish-studio.example and linked its GA4 property, Lanternfish Studio Web. Search Console has two properties for this site and I can't tell which you use: 1. sc-domain:lanternfish-studio.example (domain property) 2. https://lanternfish-studio.example/ (URL prefix) 3. Skip for now" (The owner typed 1.)
| Call, same session | next_step.code after it |
|---|---|
get_ | add_website |
add_website | choose_ |
connect_ (GA4) | importing |
connect_ (Search Console) | importing |
"Your dashboard link is below. Expect the history to land within minutes; the page updates by itself. I left Google Ads alone; say the word if you want it."
Next morning the owner pasted the status check. get_setup_status returned ready, and the agent offered to start on questions.
Notice where the agent spoke up. GA4 had exactly one unused match, so it sat in ready_to_link and was linked without a word. Search Console did not: Google lets one site exist as a domain property, covering every subdomain and protocol, and as a URL-prefix property covering only addresses with that exact prefix, per Search Console Help (as of November 2026). Both matched, so the agent asked. Picking the domain property is usually right when a site serves pages from both www and the bare domain, because it sees both.
The print shop waits. A second site moves you off the free plan, so check pricing
If your agent already acts as your AI chief of staff for the week's numbers, this run is usually the last thing standing between it and your traffic and search data.
You sign in, the agent matches the properties, and a dashboard link comes back.
When the import lands, ask your agent which landing pages turned visitors into conversions in the last 30 days.
Free for 2 sites · No credit card
Rules the agent follows, and why they exist
The guide closes with four rules. Each shuts a door an agent with write access could otherwise walk through.
It never handles passwords or Google credentials
You sign in only in your own browser, and there is no API key to create, so the agent never holds your password or a key you created. What your AI app does keep is a DuckDive access token from the sign-in, which you can revoke from either side. An agent that asks for a Google password in chat is not following DuckDive's guide.
It asks before moving a data source
If the GA4 property for your site is still attached to an old test site in DuckDive, the agent stops and asks; nothing moves until you say so. The bulk onboarding guide walks through the mechanics when it happens across a long list.
It treats your data as data
Website names, search queries, page paths, keywords, account names and reviews are written by other people, and any of them could carry text dressed up as an instruction. The guide tells the agent to read them strictly as data and never act on them; our guide to connecting AI to analytics safely explains how that kind of prompt injection works.
Only admins set things up
Setup actions need an Agency Admin or Client Admin seat. Everyone else gets a read-only connection, whatever they tick on the approval screen.
Even with actions allowed, the ceiling is low. DuckDive's connector guide names three things the connector can never do: delete data, add or remove your Google connection, or email anyone except you. To cut access, switch DuckDive off on its AI Connectors page or remove it from your AI app; DuckDive says either ends access at once.
Reading a stalled setup from its status
When a run stops short, the reason usually sits in a field the agent already read from get_setup_status or list_available_sources. Match what you saw in chat to a row.
| What you see in chat | What the status or source list reports | The fix |
|---|---|---|
| A refusal to add the website | connection_: false, on an admin seat | Finish on the setup page in your browser, or disconnect DuckDive in your app, reconnect and keep Also allow actions ticked |
| A request to ask your DuckDive admin | A team member role; next_step.code is ask_admin | An Agency Admin or Client Admin runs the prompt; your seat keeps read access meanwhile |
| A Google link and a request to type "done" | next_step.code is connect_google, typical after an email sign-up | Open the link, choose the Google account holding Analytics and Search Console, click Allow, reply "done" |
| A numbered list of properties | next_step.code is choose_sources; the source list calls one source unclear | Reply with a number, or "skip" and link it later |
| The agent passes over a property you expected | The item is flagged linked_ | The guide says to skip it; ask whoever manages the other DuckDive website |
| A source never shows up at all | Not in list_ | For GA4 or Search Console, get added in Google first; for Ads, Local Services Ads, Yext or Local Falcon, use the integrations link |
Field and code names: DuckDive agent setup guide (agents.md), read November 16, 2026.
Seats in one line: admins set up, team members read; the safety guide has the full seat and tick-box matrix. Unsure which row you are in? Let the agent read it:
Ask your agentSetup stopped halfway yesterday. Read my setup status, tell me which step it stopped at and why, and finish it if this connection is allowed to.
Tools it calls get_setup_status list_available_sources connect_website_sources
Worked example: a team member's agent that is sent to the admin
Same fictional studio, a month on. The content lead, on a team member seat, has connected DuckDive in Cursor and asks her agent to add the print shop so she can compare it with the portfolio site.
Add our print shop, lanternfish-prints.example, to DuckDive and hook up its Search Console so I can compare it with the main site.
get_setup_statusseat is a team member, next step is ask_adminlist_websitesone website visible, lanternfish-studio.example, with GA4 and Search Consoleget_sync_statusboth sources on that site are ready
"I can't add websites from your seat. You're a team member on Lanternfish Studio's DuckDive account, and setup needs an Agency Admin or Client Admin. I can read lanternfish-studio.example, where GA4 and Search Console are both ready, but adding the print shop and linking its Search Console both need your admin."
"Want me to draft a note for your admin? Something like: please add lanternfish-prints.example to DuckDive; the quickest way is to paste the setup prompt into your agent with the address in it. Meanwhile I can pull the main site's top search queries for the last 90 days."
The point is what the agent did not try: no workaround, no request for the admin's login, no second account. ask_admin is a stop sign, and the agent stayed useful with the read access the seat does have.
Skipping a source now and linking it later
You never have to link everything in one go. When list_available_sources marks a source as unclear, the agent asks once, and "skip it" is a perfectly good answer. Linking is what starts the history import, so a skipped source simply has no data until you link it.
Link a skipped Google source
Ask your agentI skipped Search Console when you set up my site last week. Find the right property and link it now, but check with me first if another site in DuckDive already uses it.
Tools it calls list_websites list_available_sources connect_website_sources
When your Google login cannot see the property
DuckDive can only offer what your connected Google logins reach, and no agent can grant Google access. Fix it in Google, then ask again:
- GA4: someone with the Administrator role on the account or property adds your Google account as a user, per Google Analytics Help (as of November 2026).
- Search Console: a property owner adds you, since only owners can add or remove users, per Search Console Help (as of November 2026).
Ads, Local Services Ads, Yext and Local Falcon
The agent leaves Google Ads and Local Services Ads alone unless you ask, and points you to the integrations link in your setup status. Yext and Local Falcon never pass through Google's sign-in, and no setup tool links them, so they wait on that integrations page too.
Ask your agentYou just linked that Search Console property. Is its import still running? Once it shows ready, tell me how recent its newest data is.
Tools it calls get_sync_status
Each source comes back from get_sync_status in one of three states (ready, importing or failing) with a note of how fresh its data is. A source that stays on failing is a question for hello@duckdive.ai, not something to unlink and relink on a hunch.
When to click through setup yourself instead
Agent setup saves clicks; it cannot fix a missing prerequisite.
- Your app cannot add remote MCP servers with OAuth sign-in. There is nothing for the agent to connect to. The remote MCP server page for other clients lists what a client needs.
- You are on ChatGPT Free or Plus. DuckDive's connector guide says ChatGPT needs Business or Enterprise. The ChatGPT setup page covers the plan question in full.
Neither route creates a GA4 property, installs a tracking tag or grants Google access; setup links data that already exists. Of the six sources DuckDive reads, an agent run links two by default: GA4 and Search Console.
The route without an agent: sign in, then open the DuckDive welcome page, which walks through adding a website, connecting Google and choosing data sources. The sign-up screen itself asks whether you want to set up with AI or by yourself, so you can choose at the door.
What people ask before handing setup to an agent
Can my AI agent sign up for DuckDive by itself?
Nearly. A terminal agent such as Claude Code can register the DuckDive server itself; in a chat app you press the app's Add or Connect button. After that the agent checks your setup, adds the website and links the GA4 and Search Console matches it is sure of. Your account is created when you select Continue with Google on the DuckDive page that opens, so there is no separate sign-up. That sign-in stays with you.
Which agents can run the setup?
Any AI app that can add a remote MCP server and sign in to it with OAuth. DuckDive's setup guide gives steps for Claude, ChatGPT, Cursor, Claude Code, VS Code, Gemini (US personal accounts), Grok and Gemini CLI; its connector guide adds that ChatGPT needs Business or Enterprise. Depending on the app, the menu may be labeled Connectors, Apps or MCP, and those labels move between versions.
Do I have to create an API key for my agent?
No. The guide tells the agent not to add an Authorization header and not to ask you for a key. Your AI app signs in with OAuth when you approve the connection, and DuckDive hands it an access token. There is nothing to paste, store or rotate.
How soon can I ask questions after setup?
Linking a source starts the history import. DuckDive puts a typical import at a few minutes, with years of history taking more. Ask your agent to check the sync status; once each source shows ready, ask your first real question.
Sources
- DuckDive agent setup guide, agents.md: setup states, agent rules and source flags. Read November 16, 2026.
- DuckDive help page for AI apps, docs/ai-connectors: ChatGPT plan note, connector limits, disconnecting. Read November 16, 2026.
- Model Context Protocol, authorization chapter of the latest spec (revision dated 2026-07-28 when read): browser approval before a token. Read November 16, 2026.
- Google, Add a website or platform property to Search Console: domain versus URL-prefix properties. Read November 16, 2026.
- Google, Add, edit, and delete Analytics users and user groups: who can grant GA4 access. Read November 16, 2026.
- Google, Managing owners, users, and permissions: who can add Search Console users. Read November 16, 2026.
Last reviewed . Lanternfish Studio and its websites in EXAMPLE DATA blocks are fictional.
