Sitebulb MCP Server: Chat With Your Crawl Data in Claude and ChatGPT

This information is correct as of August 2026. Sitebulb MCP is in beta, so details may change; Growth Method has made every attempt to ensure the accuracy and reliability of the information provided.

Is there a Sitebulb MCP server that lets Claude or ChatGPT read your crawl data? Yes. Sitebulb MCP is a hosted, read-only Model Context Protocol connector that you add to your AI assistant once, sign in to with your Sitebulb account, and then query in plain English. It works with both Sitebulb Cloud and Sitebulb Desktop, exposes 15 read-only tools covering projects, audits, Hints and URLs, and cannot start, change or delete a crawl. Sitebulb’s own summary of the idea: “Ask questions in plain English instead of jumping between reports and exports.”

That last part matters more than it sounds. Anyone who has run a technical audit knows the real work is not the crawl; it is the hours afterwards, filtering Hints, exporting URL lists, cross-referencing them with traffic, and writing the whole thing up for someone who does not know what a canonical is. Sitebulb MCP moves that stage into the place most of us already do our thinking. We are in the early access group, and this article will grow as we put the beta through its paces on real audits.

What Sitebulb MCP is (and is not)

Sitebulb is a website crawler and technical SEO auditing tool, available as a desktop app for Windows and Mac and as a cloud service. Its distinguishing feature is the “Hints” layer: rather than dumping raw crawl data on you, it flags issues and opportunities, scores their severity, and explains why they matter.

Sitebulb MCP is the bridge between those audits and your AI assistant. In Sitebulb’s own words, it brings “your projects, audits, issues and URL data into Claude, ChatGPT and other AI assistants”. It is not a new AI feature inside Sitebulb, and Sitebulb’s co-founder Patrick Hathaway was pointed about that distinction when announcing it:

“It’s about giving the assistants you already use a proper way to read your real audit data, with all the context and none of the guessing. Same crawler, same hints, same data you already trust. Just accessible from wherever you do your thinking now.”

Patrick Hathaway, co-founder, Sitebulb (LinkedIn)

If MCP itself is new to you, our introduction to MCP for marketers covers the basics. The one-line version: MCP is the open standard that lets an AI assistant call outside tools and pull their data into a conversation, so the assistant answers from your real Sitebulb account rather than its general knowledge.

How the data is organised

Sitebulb’s capabilities doc opens with a mental model that explains almost everything the MCP does. In Sitebulb:

• A project is a website you have set up to crawl.
• Each crawl of that project is an audit.
• Every audit contains reports (grouped views such as On Page or Indexability).
• Reports contain Hints (issues and opportunities, each with a severity).
• Each Hint applies to a set of URLs.

The MCP reads down that chain: project, then audit, then the reports and Hints inside it, then the URLs. Nearly every question you will ask maps to one of those levels, which is why prompts work best when they name the project as it appears in Sitebulb.

What it can do

Everything works off a crawl that has already finished. Sitebulb lists eight capabilities:

1. Find your projects and audits. List what is in your account and when each was last crawled.
2. Summarise an audit. The site score out of 100, how many URLs were crawled, and the breakdown of issues by severity.
3. Surface issues by category. Hints grouped the way Sitebulb groups them (on-page, indexability, links, security, performance and so on), each with severity, the number of URLs affected, and whether that count went up or down since the last crawl.
4. Pull informational insights. The contextual findings that are not problems as such but are worth knowing, such as the indexable URL count.
5. Show example URLs for an issue. Ask which pages are missing an H1 and get a sample to eyeball.
6. Look up a single URL. The full detail Sitebulb holds on one page: status, word count, titles, canonicals and so on.
7. Tell you what changed. Between an audit and the one before it: what is new, what is gone, which issues moved.
8. Show trends over time. Site score and key counts across past crawls, so you see the direction of travel rather than a snapshot.

Report tables also come back with a built-in steer on how they are best shown (table, bar, pie or line), so “chart the issues in my latest audit by severity” tends to render sensibly first time.

What it can’t do

The limits are deliberate and, for the most part, sensible:

• It cannot run, schedule or change a crawl. Set those up in Sitebulb as you always have; the MCP reads the results once they are done.
• It will not give you every affected URL. Each Hint returns a sample of up to 50 example URLs. For the complete list, use the export in the app.
• It cannot read a crawl that has not finished.
• It will not invent data. If it cannot find a project or audit, it says so rather than guessing.
• It cannot change your settings or your data.

That first point is the one that shapes how you use it, and it is the biggest philosophical difference from Screaming Frog’s approach (more on that below).

Cloud vs Desktop: where your data actually lives

Sitebulb MCP works with both products, but the plumbing differs, and it is worth understanding before a security team asks.

Sitebulb Cloud. Your assistant’s request goes to https://mcp.sitebulb.com/cloud, which reads the answer from your Cloud workspace and returns it to the chat. No local app is involved.

Sitebulb Desktop. Your assistant’s request goes to https://mcp.sitebulb.com/desktop, which reaches the specific Desktop machine you authorised. Your crawl data stays on that computer until a question reads it, and the hop only works while Sitebulb Desktop is installed, signed in, running, on an active licence, online, and on the latest version. If the app is closed, the assistant is told the connection is unavailable rather than being shown stale data.

That Desktop model is the clever bit. It means a crawl sitting on your laptop can be queried from the Claude web UI, with no terminal, no config file and nothing extra running beyond the Sitebulb app you already have open. It also means the laptop has to be on.

Setting it up

Setup is the same shape everywhere: add Sitebulb as a connector, sign in and approve, start prompting. The Start Here doc links a specific guide for each combination of product and client:

Taking Claude as the example, the steps are:

1. Open Claude’s connector settings (Settings, then Connectors) and choose Add custom connector. On a Team plan an Owner has to do this step.
2. Name it “Sitebulb Cloud” or “Sitebulb Desktop” and paste the matching URL: https://mcp.sitebulb.com/cloud or https://mcp.sitebulb.com/desktop. Leave the OAuth client fields blank; registration is handled automatically.
3. Click Connect. Your browser goes to auth.sitebulb.com, where you sign in with your Sitebulb credentials, choose the workspace or Desktop machine to read from, and click Allow to grant read-only access. The approval screen is worth a moment: it names the client asking (the provider domain plus the name the client registered with), the target you picked (for Cloud, your workspace address, such as yourteam.uk.sitebulb.com), and spells out what the client can do (read completed projects, audits, reports, charts, URLs, prompts and help resources) and cannot do (start, stop, edit, delete or create audits, projects, licences, users or account settings). It also warns that only one Cloud target can be active for a given client at a time, so choosing a different workspace later replaces, and revokes, the earlier one.
4. Sitebulb emails you a “connection connected” confirmation listing the provider, connector (cloud or desktop), target and timestamp, with a link to auth.sitebulb.com/connections if it was not you. Keep it; it is your record of when the connection was made and to what.
5. In a chat, open the tools menu and enable the Sitebulb connector, then test it with a simple prompt.

Two habits Sitebulb recommends from the start, and which we would second: begin prompts with “Using the Sitebulb MCP…” so the assistant pulls your data rather than answering from memory or the web, and name the project exactly as it appears in Sitebulb.

On plans: Sitebulb’s FAQ says the MCP is available on all Cloud plans and on Desktop Pro and Lite licences, and the Desktop trial is supported. Claude and ChatGPT are the formally supported clients, but because this is a standard remote MCP server with OAuth, the FAQ also lists Perplexity, Cursor and Claude Code as working. In practice, most hosted MCP servers that follow this pattern (Notion’s does the same) work anywhere that supports remote connectors. We connected Sitebulb Cloud to Growth Method that way, as a custom integration using Sign in (OAuth) and the /cloud URL, and the same approval screen and confirmation email came through with growthmethod.com as the requesting provider.

If you are on Desktop and in the beta, install the beta build from the downloads page first; per Sitebulb’s early access email, the current public release does not include the MCP yet.

Prompts worth stealing

Sitebulb has published a prompt library grouped by task. A handful that show the range:

• Get your bearings: “Summarise the latest audit for the [project] project.”
• Site health: “What’s the site score for the latest [project] audit, and what’s dragging it down?”
• Content: “How many images are missing alt text in [project], and is that better or worse than last crawl?”
• Indexability: “How many indexable pages does the latest [project] audit find, and what’s being excluded and why?”
• Change tracking: “Compare the two most recent [project] audits and tell me what changed, wins first.”
• Reporting: “Give me a three-paragraph summary of [website]‘s technical health for a marketing director.”
• Cross-tool: “Cross-reference the broken pages in the latest [project] audit with their traffic from [Ahrefs/GSC].”

That last category is where this becomes useful for marketing teams rather than just SEO specialists. The Sitebulb MCP only ever returns Sitebulb data, but your assistant can reason across every tool you have connected to it. Pair it with the GA4 MCP or a Search Console connector and “which broken pages are costing us the most visits” becomes a one-line question. Connect Jira or Linear and “turn the top five issues into tickets” is the next line.

What a first afternoon with it looks like

The prompt library is a good menu, but the useful thing to know is what a real session turns into. Here is what we did with the Sitebulb MCP in Claude Code on this site, growthmethod.com, in the first afternoon after a crawl finished. The prompts are the ones we typed; the interesting part is what the assistant did with the read-only data once it also had the codebase open.

“Which Sitebulb projects do you have access to?” A sanity check that doubles as a security check. It came back with the two projects on the workspace, their audit counts and last-crawled times. A follow-up (“is there a way to only grant access to specific projects?”) produced a straight answer: no, the OAuth grant is workspace-wide, so keep client work in a separate workspace if that matters.

“Which are the top issues from the recent crawl?” The assistant walked the chain itself: project, latest finished audit, summary (score 96, no criticals, 3 high, 9 medium), then every Hint category, then sample URLs for the actionable ones. It came back with a ranked list rather than a data dump, and separated real problems from false positives (a page that mentions “lorem ipsum” in prose is not a placeholder-text bug).

“What’s the image in number 2, GrowthLoop.jpeg? On which page?” A hotlinked diagram from a third-party site had started redirecting to a 404. Because the assistant could grep the repo, it answered with the file and line, and swapping in a locally hosted image took one more message.

“Let’s fix the 34 broken outbound links together.” This is where the read-only MCP paid for itself. For each flagged URL the assistant found the article that linked it, followed the redirect chain, searched for the content’s new home (Kevin Indig moved to Substack, ProfitWell became Paddle, Lenny’s slug changed), and either repointed the link, swapped in a Wayback snapshot for a cited-but-deleted page, or removed a dead further-reading bullet. Twenty-nine files, one commit. The next crawl showed 34 down to 3, and asking about the stragglers surfaced that we did not link them directly at all: audit_url_get returned UrlSource: Redirect, meaning an older URL we did link (blog.close.com, tractionbook.com) had started redirecting into them.

“There’s a more recent crawl.” The Lost & Found tools compare two audits, so the follow-up was a before-and-after table (external broken redirects 34 to 3, internal redirects 2 to 1, no-anchor-text pages 495 to 473) rather than a second full read.

“Only one followed internal link, something links to old slugs for those two.” Two consolidated articles had been redirected but two other articles still linked the old paths. The assistant found both links, and, because it could see the whole src/ tree, swept every redirected path in _redirects against every internal link to confirm those were the only two.

“Duplicate H1, should I fix this? Is it an issue?” Not every Hint deserves a fix, and asking for a judgement rather than an action got one: barely an SEO issue, but the /get-started/ H1 was a positioning statement rather than a description of the page, so change it for intent and clarity, not for the crawler.

“Investigate the sitewide ‘internal link with no anchor text’ hint.” The most instructive one. Raw HTML and a headless-Chrome render both showed no empty anchors, so it was a counting rule, not a bug. The Links dashboard settled it: exactly one unique internal anchor with missing text, whose destination content type was XML. That was the footer RSS icon, which had an sr-only span but no aria-label; the logo and search icons, which carry aria-label, passed. One attribute in one component, and the Hint should clear from all 473 pages on the next crawl.

Three habits fell out of the afternoon. Ask for a ranked list, then work down it one Hint per message. Ask “is this an issue?” before “fix this”, because a crawler cannot weigh a Hint against your positioning. And treat 403s on external URLs from a Cloud crawl as “verify locally”, because Sitebulb Cloud fetches from datacentre IPs that Cloudflare-protected sites will challenge, while the same URLs open fine from a desktop.

Beyond questions: reports, dashboards and Skills

Sitebulb’s What you can build guide pushes past Q&A into three things: reports you can hand to someone (plain-English summaries, client write-ups, developer lists), dashboards (single-audit overviews, score trends, wins-and-regressions boards), and workflows that pull Sitebulb together with the rest of your stack. Their own internal example is a weekly Sitebulb crawl that Claude picks up automatically, analyses, and reports on in Slack, which they describe as “essentially the ‘website monitoring’ feature we get asked to build all the time”.

The more interesting layer is Sitebulb Skills. If the MCP provides the data, a Skill provides the process: a packaged job with the questions to ask, the order to read the data, and the deliverable to produce, so you run it in a sentence rather than a long prompt. Sitebulb ships five, installed together as a plugin from the sitebulb/sitebulb-plugin repository, in Claude, ChatGPT desktop or Claude Code:

• What To Fix First: triage for your first five minutes with a crawl. Asks about the site and situation, then ranks issues by impact against effort.
• What Changed: compares the two most recent crawls into wins, regressions and a watch-list. Built for the monthly retainer conversation.
• Technical SEO Audit: turns a crawl into a client-facing audit document, delivered as a draft with review comments where your judgement is needed.
• Release Check: a smoke alarm for teams shipping site changes on a cadence. Set a watch-list once and get a two-line “all clear” (or not) after each scheduled crawl.
• Dev Handoff: writes tickets into Jira, Linear, Asana, Monday or ClickUp, each with the problem, why it matters, sample URLs and an acceptance criterion, previewed before anything is created.

Two of those (What Changed and Release Check) are designed to run automatically the moment a scheduled crawl lands, using something that can run on a schedule such as Claude Cowork. The constraint, again, is that the MCP cannot start the crawl, so automation always means the crawl is already scheduled in Sitebulb. For a broader take on when to reach for an MCP, a CLI or a Skill, see MCPs, CLIs and Skills: which fits your workflow.

Dragan Berak of Rock The Rankings, writing on Sitebulb’s blog about running SEO and AEO audits inside Claude with GSC, GA4, Ahrefs and Sitebulb data in one context, put the value plainly:

“This isn’t about letting AI write the audit for you. It’s about eliminating the friction between data collection, analysis, and writing.”

Dragan Berak, Rock The Rankings (Sitebulb)

His guide predates the MCP and relies on pasting an “All Hints” export into Claude. The MCP removes that step.

Sitebulb MCP vs Screaming Frog MCP

If you follow this space you will know Screaming Frog shipped a built-in MCP server in v24, and we have written about running Screaming Frog in the cloud and driving it through MCP. The two tools have landed on almost opposite designs, and the difference is worth understanding before you pick.

Neither is right in the abstract. Screaming Frog’s model is the more powerful one for an agentic setup where you want the assistant to own the whole loop, and it is what makes a cloud crawler your whole team can talk to possible. Sitebulb’s model is the easier one to approve: nothing the assistant can be prompted to do will crawl a site, hammer a server or delete an audit, which is the guarantee a security reviewer or an agency’s client will ask for. Plenty of teams will end up with both.

Security: what a reviewer will want to know

Sitebulb’s security and privacy overview is unusually straight for a vendor doc, and the details are the ones an IT review usually needs:

• Read-only, by design. The MCP exposes 15 tools and every one of them reads. Sitebulb is explicit that “this isn’t a restriction we’ll loosen quietly later; it’s how the connector is designed”.
• OAuth 2.0, single scope. Sign-in happens at https://auth.sitebulb.com, never in the AI client, and requests one permission: mcp:read. Your Sitebulb password is never shared with the assistant.
• No API keys. The sign-in flow is the only way to connect, so there is no long-lived secret for anyone to leak.
• Per-user connections. Each person signs in as themselves and the assistant sees only what their own Sitebulb account can see. One person’s access can be revoked without touching anyone else’s.
• You are told when a connection is made. Every approval triggers a confirmation email naming the client, connector, target and time, with a link to auth.sitebulb.com/connections to revoke it if it was not expected. That is a small thing, but it is the audit trail most connectors do not give you.
• One Cloud target per client. A given client can hold one active Cloud workspace connection at a time; picking another replaces and revokes the first, so a client cannot quietly accumulate access to several workspaces.
• Nothing syncs in the background. Data moves only when a question needs it, and only the data that question needs.
• Blast radius. Worst case, a connected assistant can read what the connecting user’s account can read. It cannot modify, delete or crawl anything.
• Certification. Sitebulb does not currently hold SOC 2 or ISO 27001.

The part Sitebulb does not control is what happens once a result lands in your chat. From that point it is handled by Anthropic or OpenAI under your agreement and settings with them, including whether it is used for training. If your organisation has rules about what goes into AI tools, those rules apply to Sitebulb data from the moment it enters a conversation. And revoking access, which takes under a minute by removing the connector from the AI client or from the connections page the confirmation email links to, stops future reads but cannot reach back into chats that already happened.

Beta caveats

This is a beta, and Sitebulb says so. Expect rough edges, and if you hit a bug the team has asked for feedback by replying to the early access email. A few things from the troubleshooting FAQ that will save you a support ticket:

• On Desktop, the app must be open, signed in, updated, licensed and online.
• The MCP only reads finished audits, so a “can’t find that audit” is usually a crawl still running or a project name that does not match exactly.
• If tools are missing from a chat, enable Sitebulb in the chat’s tools control, and hard-refresh if needed.
• Rate limiting is normal protective behaviour; wait a moment and retry.
• If the assistant answers from the web instead of your data, remind it: “Use the Sitebulb MCP for this.”

Getting started checklist

1. Join the waitlist if you are not already in the beta. Cloud workspaces are updated automatically; Desktop users install the beta build from the downloads page.
2. Make sure at least one project has a finished crawl.
3. Follow the Start Here guide for your product and client, using https://mcp.sitebulb.com/cloud or https://mcp.sitebulb.com/desktop.
4. Run the three-prompt warm-up from the getting started guide: show me my projects, summarise my latest audit, what would you fix first.
5. Install the Skills plugin and run What To Fix First on a real project.
6. Connect the rest of your stack (GA4, Search Console, Jira, Slack) to the same assistant and start asking cross-tool questions.

About Growth Method

Growth Method is the agentic marketing platform for B2B teams. Pre-built AI agents plan, launch and analyse marketing campaigns from your live marketing data, with your martech stack connected in minutes. Crawl data is one input among many: an audit finding becomes a campaign, gets prioritised against your goal, and is measured on the business outcome rather than the Hint count. If you would rather run more campaigns than manage exports, get started here.

Further reading

Sitebulb’s own documentation, in the order you will probably need it:

• Sitebulb MCP: Start Here
• Getting Started with the Sitebulb MCP
• What the Sitebulb MCP can and can’t do
• Sitebulb MCP prompt library
• What you can build with Sitebulb MCP
• Sitebulb Skills for Claude and ChatGPT
• Sitebulb MCP: security and privacy overview
• Sitebulb MCP: troubleshooting and FAQ
• Disconnecting and revoking Sitebulb MCP access

Related reading from us:

• How to run Screaming Frog in the cloud (and drive it through MCP)
• MCPs, CLIs and Skills: which AI interface fits your marketing workflow?
• GA4 MCP server: connect Google Analytics to Claude and Cursor
• An introduction to MCP for marketers

Frequently asked questions

What is the Sitebulb MCP server?

Sitebulb MCP is a hosted Model Context Protocol connector from Sitebulb that lets AI assistants such as Claude and ChatGPT read your Sitebulb projects, audits, Hints and URL data. You add it as a connector once, sign in with your Sitebulb account, and then ask questions in plain English instead of exporting reports. It works with both Sitebulb Cloud and Sitebulb Desktop.

Is the Sitebulb MCP read-only?

Yes, by design. Sitebulb says the MCP exposes 15 tools and every one of them reads. There is no tool that can start, pause, stop or delete a crawl, change a setting, or write anything back into Sitebulb. Crawls are set up and scheduled in the Sitebulb app as usual; the MCP reads the results once they finish.

Does Sitebulb MCP work with Sitebulb Desktop or only Sitebulb Cloud?

Both. Cloud connections go to https://mcp.sitebulb.com/cloud and read from your workspace. Desktop connections go to https://mcp.sitebulb.com/desktop and reach the specific machine you authorised, so Sitebulb Desktop must be installed, signed in, running, on an active licence, online, and on the latest version for the connection to work. If the app is closed, the assistant is told the connection is unavailable.

Which AI assistants work with Sitebulb MCP?

Claude and ChatGPT are formally supported at launch, on both web and desktop, and Sitebulb publishes separate setup guides for individual and team/business plans on each. Because it is a standard remote MCP server with OAuth sign-in, Sitebulb’s FAQ notes it also works in other MCP-capable clients such as Perplexity, Cursor and Claude Code.

Do I need an API key to use Sitebulb MCP?

No. There are no API keys. You connect through an OAuth 2.0 sign-in at auth.sitebulb.com with a single read-only scope (mcp:read). Each connection is tied to an individual user, confirmed by email when it is made, and can be revoked by removing the connector from your AI client or from the connections page at auth.sitebulb.com/connections.

How is Sitebulb MCP different from the Screaming Frog MCP server?

They take opposite approaches. Screaming Frog’s MCP server runs locally alongside the Spider (or on a cloud machine you control) and can drive the crawler: start crawls, poll progress, pull reports and run scripts. Sitebulb’s is a hosted, sign-in connector that is strictly read-only and only reads finished audits. Screaming Frog suits teams who want an agent to operate the crawler; Sitebulb suits teams who want the crawler to stay under human control and the assistant to do the analysis and reporting. Our Screaming Frog cloud guide covers the former in detail.

Can Sitebulb MCP give me every affected URL for an issue?

No. For any single Hint it returns a sample of up to 50 example URLs to keep answers fast and readable. When you need the complete list, use the export in the Sitebulb app.