A Screaming Frog crawl means thirty tabs, dozens of filters and just as many exports to dig out. The useful work, deciding what to fix first, starts after that. The MCP server moves that line: Claude launches the crawl, opens yesterday's one, pulls the right export, and can even write a Node script to process the file before handing it back to you.
I installed it live in the video below (in French), on a brand-new machine. Allow five minutes, three settings, and one trap that catches everyone on the first test.
Screaming Frog × Claude: the full setup in 3 min 45 s (in French)
Video chapters (in French): 0:16 database storage · 0:49 enabling the MCP server · 1:35 the extension in Claude · 2:37 the test and the trap.
Key takeaways
- The Screaming Frog MCP server is reserved for paid licenses, on version 24 or later.
- Three settings are enough: database storage, the Node.js environment accepted, and the extension installed in Claude Desktop.
- Two modes to choose from: STDIO, where Claude launches Screaming Frog in the background, or Streamable HTTP, where the interface stays in front of you.
- The first-test trap: as long as the crawl is still open in the interface, Claude cannot find it. Clear it, and it appears.
- On the Claude side: Claude Desktop through the official extension, Claude Code through the local address of the HTTP mode. Claude on the web cannot reach a server installed on your computer.
- The MCP server saves you the time spent on exports, not the time spent on analysis: prioritizing is still your job.
How to connect Screaming Frog to Claude: the short answer
- Check that you have a paid license for Screaming Frog SEO Spider, on version 24 or later.
- Switch to database storage:
File > Settings > Storage Mode. - In
File > Settings > MCP Server, accept the Node.js environment and choose the working folder. - Download the
.mcpbextension from the official guide and install it in Claude Desktop:Settings > Extensions > Advanced Settings > Install Extension. - Fully quit Claude, reopen it, then ask for a crawl in plain language.
Do you work in Claude Code rather than in the desktop app? The procedure is further down, in Connect Screaming Frog to Claude Code. Claude on the web, for its part, cannot connect to Screaming Frog: it reaches MCP servers from Anthropic's infrastructure, whereas the Screaming Frog server runs on your machine.
What the Screaming Frog MCP server actually changes
MCP, short for Model Context Protocol, is a standard plug: on one side, a piece of software exposes what it can do; on the other, an AI model plugs into it and calls those functions itself. Here, Screaming Frog plays the role of the server and Claude the role of the client. If the concept is new to you, I explain in detail what an MCP server is and how it works.
Once it is plugged in, Claude gets access to three families of tools. First, the crawl and export tools: start a crawl, follow its progress, reopen a saved crawl, pull a tab, a filter or a full report. Next, a Node.js engine, which lets it write a script to process an export that is too large and install the npm packages it needs. Finally, file access, limited to a folder you choose, to read and write the results.
That autonomy is what makes the difference with a simple manual export: you describe the deliverable, and Claude chains the technical steps to get there.
What the setup gives you
Three settings to make once, and the crawl can be queried in plain language. Here is what it changes day to day, and what it does not.
- Exports and reports requested in one sentence
- Two crawls compared without opening a spreadsheet
- Node scripts generated on the fly on your data
- Paid Screaming Frog license required
- The context window fills up very quickly
Setup time: 5 minutes Screaming Frog 24+ · Claude Desktop · macOS or Windows
What you need before you start
- A paid Screaming Frog license, on version 24 or later. The free version does not expose the MCP server: the setting does not even appear.
- Claude Desktop, the desktop app: the extension cannot be installed from the browser. LM Studio also works, with a local model, and so does Claude Code through the HTTP mode described below.
- Database storage enabled in Screaming Frog. It is the default mode in version 24, but it is worth checking.
Where to find the official procedure
Everything is documented in the official Screaming Frog MCP server guide, with the two extensions to download, the LM Studio configuration and the full list of tools. It is the procedure I follow in the video.
Step 1: switch to database storage
Go to File > Settings > Storage Mode, and choose database storage rather than memory (RAM) storage. The application then restarts.
This setting is not cosmetic. In memory mode, a crawl disappears when you close the tab: it exists nowhere on disk, so the MCP server has nothing to reopen. In database mode, every crawl is saved, has an ID, and can still be looked up weeks later. That is what lets you ask "compare the site's two latest crawls" without preparing anything.
Step 2: enable MCP in Screaming Frog
Still in the settings: File > Settings > MCP Server. Two checkboxes matter there.
Accept the Node.js environment. This is what allows the model to run scripts and install npm packages on demand. Without it, you keep the exports but lose all the processing, and that is precisely what avoids pushing 50,000 rows through the conversation. The same screen lets you choose the folder where scripts, exports and installed packages land: pick a dedicated folder, not your desktop.
Start the MCP server. This checkbox only matters in Streamable HTTP mode, described below. In STDIO mode, Claude takes care of launching Screaming Frog on its own.
Step 3: install the extension in Claude
From the user guide, download the extension for the mode you chose, a .mcpb file. Then, in Claude Desktop: Settings > Extensions > Advanced Settings > Install Extension, select the file and confirm.
Two details will cost you a quarter of an hour if you miss them:
- The extension assumes Screaming Frog is installed in the default location (
/Applicationson macOS,C:\Program Files (x86)on Windows). Anywhere else, you need to edit the file to fix the path. - You must quit Claude completely, not just close the window. On Windows, the app keeps running in the system tray: right-click the icon, "Quit". Otherwise the extension will not be loaded on restart.
Reopen Claude: in Settings > Developer, the extension should appear and show that it is running. From then on, you talk to it normally.
Step 4: the first crawl, and the crawl-clearing trap
The test I run in the video fits in one sentence:
Crawl https://example.com with Screaming Frog, then summarize the main issues found, in order of priority.
Claude triggers the crawl, follows its progress, then fetches the issues overview report. Nothing to click.
The trap: clear the crawl once it is finished
As long as the crawl stays loaded in the Screaming Frog interface, it is not stored in the database: Claude does not see it in the list of recent crawls and tells you it does not exist. Clear the crawl once it is finished, ask for the list again, and it appears. It is the only real friction point in the setup.
To point to a specific crawl, you can also give Claude its ID: right-click the crawl in File > Crawls. Useful when several versions of the same site coexist.
STDIO or Streamable HTTP: which one to choose
Both modes give access to the same tools. What changes is who launches the software and what you see while it runs.
| STDIO | Streamable HTTP | |
|---|---|---|
| Who starts Screaming Frog | Claude, from the command line | You, as usual |
| Visible interface | No, everything runs in the background | Yes, you follow the data on screen |
| To do before chatting | Nothing | Start the server from the MCP menu |
| Server address | None, Claude launches the process | http://localhost:11435/mcp by default |
| Node.js to install on the machine | No, it is built in | Yes for the Claude Desktop extension, which goes through a Node proxy |
| Good for | Chaining tasks without supervision | Learning, checking, staying in control |
To get started, I recommend Streamable: watching the tab open while Claude works teaches you a lot about what it is really doing. Once you trust it, STDIO is more comfortable, and it is the one in the video.
Connect Screaming Frog to Claude Code
Screaming Frog documents two clients: Claude Desktop and LM Studio. Claude Code has no official procedure, but it accepts any MCP server over HTTP, and Streamable mode exposes precisely a standard local address. Here is how:
- Open Screaming Frog normally, with the settings from steps 1 and 2, then start the server from the
MCPmenu. - In a terminal, declare this server to Claude Code. The default port is 11435, the one the official guide gives for LM Studio:
claude mcp add --transport http --scope user screaming-frog http://localhost:11435/mcp - Launch Claude Code and type
/mcp: the server should show as connected. If not, check that it is started in Screaming Frog and that the port has not been changed.
Since this procedure is not documented by the vendor, keep an eye on its release notes. Two more remarks: by default, Claude Code caps an MCP tool's response at 25,000 tokens, one more reason to have large exports written to a file. And if the tool is not on your machine yet, my tutorial to install Claude Code walks through every step.
What Claude can do once connected
The server exposes around thirty tools (29 in the current documentation). In practice, they fall into five families.
| Family | What you can ask |
|---|---|
| Drive a crawl | Launch a crawl on a URL, follow its progress, pause it, resume it |
| Reopen existing crawls | List recent crawls, load one by its ID, compare two versions |
| Export | A tab, a filter, chosen fields, bulk exports, reports (redirect chains, issues, etc.) |
| Process | Write a Node.js script on an export, install the packages it needs, produce a table or an HTML page |
| Inspect a URL | Content, inlinks and outlinks, technical information, a screenshot of the page |
Five requests that work from day one
Prompts that get good results have one thing in common: they name the tab, the filter and the deliverable. The model no longer has to guess where to look.
- The starting point: "Crawl example.com, then list the issues from the overview report by priority, separating what is truly urgent from the rest."
- Broken links: "Export all non-200 response codes with their inlinks into a single CSV file."
- Monitoring: "Open the two latest crawls of example.com, compare the main issues and summarize what changed."
- Internal linking: "Export the HTML URLs from the Internal tab to a file, with the Address and Unique Inlinks fields, then show the top 20 most linked pages as an HTML chart."
- Images: "Export all images over 100 KB and generate an HTML page that displays them in a grid, with their file size under each one."
The principle is the same as for any request to a model: the wording is not what matters, it is the context you provide, which is the core argument of my article on prompt engineering. Here, the context is the crawl, and you still have to say which one.
The limits you need to know
The context window. This is the real constraint. An export of 40,000 URLs does not fit in a conversation, and the model chokes long before that. Hence the rule: go through files, let Claude write a script to process them, and open a fresh conversation for each task rather than stacking topics in the same one.
Security. Allowing a model to run scripts and write files is not trivial. Screaming Frog restricts access to a folder you designate, but the rule of caution is the same as for any agent: a dedicated folder, and an eye on what runs. It is the same trade-off as for any serious use of Claude, which I assess in my Claude AI review: the autonomy you gain is paid for in vigilance. One last point, a contractual one: the Screaming Frog documentation points out that no one other than the license holder should access the MCP server. So no server shared with the whole team on a single license.
What it does not replace. Screaming Frog says so itself in its documentation, and I am happy to repeat it: AI makes tasks more efficient, it does not replace analysis. Sorting 200 issues by "priority" according to a generic scale has nothing to do with deciding what matters for this site, this business, this quarter. The method does not change, whether it is my SEO audit report example or the checklist for a technical SEO audit: only the time spent building the tables goes down.
What about Search Console?
It is the question that ends the video, and it is where the setup really pays off. If you connect the APIs in Screaming Frog (Search Console, Analytics, PageSpeed, Ahrefs) before running the crawl, their columns are part of the crawl. Claude reads them like the others.
You can then ask for cross-analyses that usually take half a day in a spreadsheet: pages losing clicks that have few internal links, error URLs that were still getting traffic last month, slow pages among the ones that bring in the most. That is exactly the kind of task I delegate, and it is what turns a crawl into an action plan.
If this kind of connection interests you beyond Screaming Frog, that is what AI automation on your own tools is all about: the same plug exists for your CRM, your CMS or your accounting exports.
Frequently asked questions
How do I connect Screaming Frog to Claude?
You need a paid license on version 24 or later, and three settings. In Screaming Frog, "File > Settings > Storage Mode" must be set to database storage, and "File > Settings > MCP Server" must accept the Node.js environment. In Claude Desktop, install the spider-stdio-mcp.mcpb extension (or spider-streamable-mcp.mcpb) downloaded from the official guide, through "Settings > Extensions > Advanced Settings > Install Extension", then fully quit Claude and reopen it.
Does the MCP server work with the free version of Screaming Frog?
No. The MCP server is reserved for paid licenses, on version 24 or later. The free version, limited to 500 URLs, does not expose the server: the setting does not appear in the settings.
How do I use Screaming Frog with Claude?
Once the extension is installed, you talk to Claude normally. "Crawl example.com and summarize the issues by priority" is enough to trigger a crawl. Claude can also list recent crawls, load one by its ID, produce an export or a report, write the result to a file and generate a Node.js script to process that file.
Why can't Claude find my crawl?
Because it is still loaded in the interface. As long as the crawl is on the Screaming Frog screen, it is not stored in the database and does not appear in the list of recent crawls. Clear it once it is finished, then ask Claude to list the crawls again: it sees it right away.
STDIO or Streamable HTTP: what is the difference?
In STDIO mode, Claude launches Screaming Frog itself from the command line: the interface stays hidden, nothing to start by hand. In Streamable HTTP mode, you open Screaming Frog normally and start the MCP server from the menu: you see on screen what Claude is working on. STDIO for automation, Streamable to keep an eye on the data.
Can I connect Screaming Frog to Claude Code or to Claude on the web?
With Claude Code, yes: start the server in Streamable HTTP mode in Screaming Frog, then declare it with claude mcp add --transport http screaming-frog http://localhost:11435/mcp. This route is not documented by the vendor, which only mentions Claude Desktop and LM Studio. Claude on the web, however, cannot reach a server running on your computer.
How do I export crawl data with the MCP server?
By asking for it, specifying the tab, the filter, the fields and the file name. "Export the HTML URLs from the Internal tab to internal.csv, only the Address and Status Code fields" produces the file in the server's allowed folder. Using a file rather than the conversation keeps the context window from filling up.
Can I bring Search Console data into a crawl driven by Claude?
Yes, as long as you connect the APIs in Screaming Frog before running the crawl. The Search Console, Analytics, PageSpeed or Ahrefs columns are then part of the crawl, and Claude reads them like the others. That is what lets you prioritize fixes by the clicks actually lost.
Connect AI to your real tools, not to demos
Your crawls, your exports, your client files: I design automations that hold up in production, and I train the team that will use them.
Written by
Anthony Courtin
Reader reviews
Did this article help you? Say so, it helps the next readers.
What readers say
Be the first to leave a review.
