.theseoINSTITUTE FOR VISIBILITY
About this choice

We only save your display preference in a preference cookie, for up to 180 days on this device. No CRM or advertising data. Your tracking consent stays unchanged. If cookies are blocked, your choice may not be saved.

Discover
Close

MCP · IN SHORT

Your AI.Useful tools.Within reach.

Tools inside your assistant.Measurements and knowledge to draw on.Free to connect, no account.

Connect your AI appExplore what matters
Illustration of a titanium connection hub with lavender cables and three smoked-glass modules.
VISUAL IMPRESSION · MCP

START HERE

Choose the right connection.

One URL for everything: without an account you get thirteen free tools, with an account your personal features as well.

Personal MCP

Use this URL with your TheSEO account. Your AI app starts sign-in and asks for consent before using personal features.

You can also select and copy the URL yourself.

Open your MCP account

Free tools · no account

The same URL without an account: thirteen free tools, no key. AI visibility per model, an AI crawler check, trade rules with sources and open data.

You can also select and copy the URL yourself.

The free tools store nothing about who uses them. Read the explanation for what each tool does and does not do.

Read the connection instructions
  1. Open an AI app that supports remote MCP servers.
  2. Add the chosen server URL. For personal features, follow sign-in and the consent screen.
  3. Check which tools your app offers. Copying a URL alone does not create a connection.

WHAT MATTERS

Three things to know.

ToolsFrom a question to a practical task.

The server gives your AI access to specific tools, such as the AI crawler check, the trade rules and a search for a suitable skill.

SourcesGive the answer a basis.

The tools use our catalogue, our own measurements and open data. The new tools name their source, and you can read the explanations yourself.

OpennessNo account needed.

You need no account or key for the free tools. The code of our first server is open source under the MIT licence.

Connect once.Use it with a purpose.

You keep working in your own AI assistant. The server adds the tools.

View all details
  1. Choose your environment.

    Open the instructions for Claude, Claude Code or another compatible MCP client.

  2. Add the server.

    Copy the public server settings, following the setup instructions.

  3. Give it a task.

    Ask your assistant for a measurement, a skill or information from the available sources.

Connect the tools.

Thirteen tools, available without an account or key.

Gianluca, founder of TheSEO

A question?
A direct line.

Working on websites and visibility since 2017. You speak directly with Gianluca.

Get in touch
All information

Explore the details at your own pace.

~/mcp/mcp.theseo.nl[ok]
INS.MCP · free, with or without an account reg. M.000 · mcp.theseo.nl · 42

An MCP server with thirteen free tools.

Add mcp.theseo.nl once to Claude, Claude Code, Cursor or ChatGPT, and your AI can measure how often AI models mention your business, check whether your site blocks AI crawlers, look up the rules for a profession with their source and query open data from Statistics Netherlands (CBS), PDOK and the Dutch vehicle register. Those thirteen tools work without an account and without a key. With an account, Jarvis or the premium skills you also get memory, tasks, messages between your AIs and your own documents.

What the open base server does and does not record

The MCP server does not store conversations. For each valid tool call, the existing usage counter receives the tool name, UTC day and the number 1. When the separate outcome measurement is configured, it also receives a fixed schema version, whether the call started or finished, its status as pending, successful or failed, and a broad duration band. The bands are under 1 second, 1 to under 5 seconds, 5 to under 30 seconds, 30 seconds or more, or not measured. The public status route reported on 1 October 2026 that the existing counter was active; that observation does not prove stored points or active outcome measurement. These measurement points contain no tool input, response text, error text, IP address, headers, account data, conversation content or user identifier. We do not link them to people or individual conversations. Cloudflare Analytics Engine automatically adds a technical timestamp and, according to its documentation, retains the points for three months. We use the measurement for call counts and technical reliability, not for unique users or commercial results. Missing data and sampling are identified as limitations. A configured binding alone does not prove that points have been stored. Apart from that counter, five of the seven tools do not pass tool input to the site flow and only read from our own catalogue and open data. Two tools do pass data to the site flow: seo_scan stores the scanned web address and result; lokale_vindbaarheid_check stores the business name, location and result. This is the same recording as for the site forms on theseo.nl. What we do with it is set out in the privacy statement NL.

thirteen tools without an account, 42 with one
THE ADDRESS · THIS IS ALL YOU NEED https://mcp.theseo.nl/mcp streamable http · sign-in only for the account tools · never a sign-in: https://mcp.theseo.nl/open
~/mcp/mcp.theseo.nl · without an account
theseo ~ %
# 13 tools, no account, no key
ai_zichtbaarheid_merk ai_index_gat how often AI names you
ai_crawler_check robots.txt, llms.txt and markup
vakregels rules per profession, with source
cbs_cijfers pdok_adres kenteken_info open data
social_vermeldingen Bluesky and Mastodon
ai_index_branche ai_index_vragen ai_index_bedrijf the AI Index
skill_zoeken bedrijfsbrein_template 149 skills and the template
[ok] 13 tools loaded · 42 with an account
~/mcp/mcp.theseo.nl/00[ok]
Section 00 · What you get

One address, four layers

reg. M.000b

There are 42 tools at https://mcp.theseo.nl/mcp. Which ones your AI sees depends on what you have: thirteen without an account, more with an account, Jarvis or the premium skills. You sign in through your browser; your password never reaches your AI. If you want an address that never asks you to sign in, use https://mcp.theseo.nl/open: it only has the tools on TheSEO's own data and is intended for Claude's connector directory.

FIG.00: The four layers of mcp.theseo.nl[#----]
layers.json13 without an account · 42 in total
Without an account13 tools, ready to use TOOLSai_zichtbaarheid_merk ai_index_gat ai_crawler_check vakregels social_vermeldingen kenteken_info cbs_cijfers pdok_adres ai_index_branche ai_index_vragen ai_index_bedrijf skill_zoeken bedrijfsbrein_template WHAT YOU DO WITH THEMMeasure how often AI models mention your business and in which questions competitors are named and you are not, check your site for AI crawlers and llms.txt, Dutch rules and quality marks per profession with source, open data from CBS, PDOK and the vehicle register, public mentions on Bluesky and Mastodon, replay the AI Index, and search 149 free skills: the 100 from the skill library and 49 from articles.
With a free accountyour data and measurements TOOLStseo_mijn_gegevens tseo_toestemming tseo_vergeet tseo_meting_starten tseo_meting_uitslag tseo_meting_wissen WHAT YOU DO WITH THEMSee what TheSEO stores about you, view or withdraw a consent and erase data. Start your own measurement of how often OpenAI, Gemini and Claude mention your business, with the result and the difference from your previous measurement. A measurement is deducted from your account's credit.
With Jarvis Start or Soloyour own brain and the AI team space TOOLSjarvis_documenten jarvis_document_lees jarvis_document_schrijf jarvis_historie jarvis_document_beheer and the team space: tseo_onthoud tseo_herinner tseo_profiel_lees tseo_profiel_zet tseo_logboek_schrijf tseo_logboek_lees tseo_taak_maak tseo_taken tseo_taak_klaar tseo_bericht_stuur tseo_inbox tseo_reserveer WHAT YOU DO WITH THEMYour AI reads and saves documents in your own Jarvis brain, retrieves an earlier version and can restore a deleted document. In the team space ChatGPT, Claude and Codex share the same memory, the same tasks and the same log, send each other messages and reserve work so that nobody works on the same thing at the same time. Start costs 9 euros a month with two AI connections, Solo 29 euros with three, including VAT.
With the premium skillsskills that remember and check TOOLStseo_start tseo_skill_module tseo_toets tseo_bewijs_leg tseo_bewijs_toon tseo_verbruik, plus the same team space as with Jarvis WHAT YOU DO WITH THEMThe premium skills load their modules from the server, check a text against fixed rules and record what an AI claimed and what the server could verify itself. You can see how much you used today: the limit is 2,000 calls a day and 100 MB of storage in your own space.
FIG.00B: Connecting, the same address everywhere[#----]
CLAUDE AND THE DESKTOP APP
  1. Go to Settings and then Connectors.
  2. Choose Add connector and paste https://mcp.theseo.nl/mcp
  3. For the account tools, Claude signs you in through your browser.
CLAUDE CODE
  1. Run in your terminal: claude mcp add --transport http theseo https://mcp.theseo.nl/mcp
  2. The tools are ready in your next session.
CURSOR
  1. Open ~/.cursor/mcp.json.
  2. Under mcpServers, add a key theseo with url set to https://mcp.theseo.nl/mcp.
CHATGPT
  1. Turn on developer mode in ChatGPT.
  2. Add a custom MCP server under Connectors with the same URL.

Everything below is about the open base: theseo-mcp, the first server with seven tools, open source under the MIT licence and usable without an account at its own address. The two tools that pass data on to the form on theseo.nl, seo_scan and lokale_vindbaarheid_check, only exist on that open base; on mcp.theseo.nl they are switched off.

~/mcp/theseo-mcp/01-demonstration[run]
Section 01 · The demonstration

You ask something, the AI measures it

reg. M.001

This is what it looks like when the server is running. You type an ordinary sentence, your AI recognises that there is a tool for it, calls it with the input specified in the code, and receives measured text instead of a guess. The numbers below are deliberately shown as placeholders, because they come from your own site and not from our imagination.

FIG.01: A tool call and the response[#----] sheet 1/5
~/work · claude + theseo-mcp
you > run an SEO scan on https://example.com and summarise the biggest problems
tool callchosen by the AI itself seo_scan · { "url": "https://example.com" }
# the only input field: url, text, maximum 400 characters
theseo-mcp · request forwarded to the site Worker...
[technical]HTTPS, HTTP status, server speed, viewport, language, robots.txt, sitemap, security headers, email authentication
[on-page]title, meta description, H1, heading structure, alt text, canonical, internal links
[content]word count, content structure, indexability
[ai]structured data, FAQ schema, Open Graph, Twitter Card, entity clarity
response to the AIschematic, the numbers are from your site SEO scan of https://example.com (timestamp) Overall score: [score]/100 Technical [score]/100 · On-page [score]/100 · Content [score]/100 · AI [score]/100 [OK] Canonical URL: [the finding on your page] [PLEASE NOTE] Meta description: [the finding on your page] [CRITICAL] H1 heading: [the finding on your page]
[####] 4/4 categories ready · the AI summarises, you can see where it came from
7M.01 · tools on the server
24M.02 · points in the SEO scan
100M.03 · skills in the search tool
0M.04 · euros, MIT licence, no account
M.01 to M.04 · properties of the server as it is currently running, not usage figures or results.
$ cat 01-what-is-mcp.md
DOC.01

What MCP is, in plain language

MCP stands for Model Context Protocol. It is an open standard that defines how an AI assistant may use external tools. That sounds more technical than it is. On its own, a language model can do only one thing: produce text based on what it has read before. It cannot open your website, look in Google Maps or search a catalogue. If you ask such a model how your site is performing, you will get a plausible account based on nothing. That exact problem is why MCP exists.

The standard solves this with a simple division of roles

On one side is the client: the program in which you talk to the AI, such as Claude, Claude Code or Cursor. On the other side is a server that can do something the model itself cannot. That server tells the client which tools it offers, which input fields each tool expects and roughly what it returns. Your assistant reads that description and decides for itself when to call a tool. You do not have to learn a command or look for a button: you ask your question in plain language and the assistant chooses the tool.

The comparison that is easiest to remember: MCP is to AI tools what USB is to peripherals

Before USB existed, every device had its own plug and its own driver. Now everything fits into the same port. Without MCP, every AI environment would have to build its own connection to every service. With MCP, you write one server, and it works everywhere the standard is followed. The standard is publicly documented at modelcontextprotocol.io and most major AI environments now support it.

The practical difference is best illustrated by asking the same question twice

Without a server, you get an answer that sounds like knowledge. With a server, you get an answer you can verify.

--- without-mcp.txt +++ with-mcp.txt  # the same question, a different kind of answer "How does my site perform technically?" > the AI reasons from general knowledge no measurement, no source, no way to check it "How does my site perform technically?" > the AI calls seo_scan 24 measured points returned, each with a status and explanation, and you can check them
Important detail: the AI decides for itself whether to use a tool

You can explicitly ask it to do so, but in practice it uses the tool automatically as soon as your question calls for it. Five tools read bundled files and are registered as read-only. The SEO scan and local check are not registered as read-only: the site flow may record their input and result as a lead in the CRM. They do not modify your site, profile or source data. The server does count one call per tool and UTC day, and the scan and local check send their input to the site flow exactly as stated below.

The question that goes with this, and which is rarely answered anywhere: how does such a tool actually end up in your conversation?

That happens in two entirely separate parts. Once, when you paste the endpoint, the client and server perform a short handshake in which the server lists its entire set of tools. Then, on every turn, the model itself decides whether there is anything among them that matches your sentence. The figure below places the two side by side, because they are often confused.

FIG.01b: How a tool ends up in your conversationa one-off handshake, then a choice on every turn
connection.logschematic example · the message names come from the MCP standard
PHASE 1 · ONCE, AS SOON AS YOU HAVE PASTED THE ENDPOINT
client →initializeWhich version of the standard do you use, and what do you support?
← servertheseo-mcp 0.1.0Version confirmed. Your client sends technical MCP handshake data here, such as the protocol version, capabilities, client name and version. No account, key, site or conversation data is included.
client →tools/listWhat can you do, and what input do you need for it?
← serverseven tools, each with a name, explanation and input schemaseo_scan · lokale_vindbaarheid_check · skill_zoeken · bedrijfsbrein_template · ai_index_branche · ai_index_vragen · ai_index_bedrijf
From that point on, those seven descriptions are in your assistant's toolbox. It did not look them up because you asked, it received them when it connected. That is precisely why you do not need to learn a command anywhere: the explanation sent by the server is exactly what the model needs to make its own choice.
PHASE 2 · EVERY TURN, WITHOUT YOU HAVING TO DO ANYTHING
you type an ordinary sentence
a description matches the call is checked against the schema The model fills in the input fields, which are first checked against the schema. A URL longer than 400 characters or an empty company name is rejected before anything is sent. If it passes, you receive measured text such as that in FIG.01.
no description matches the server remains silent That is not a fault, but the intention. A question about your visual identity has no use for an SEO scan, so the model simply answers it itself. A server that called something for every sentence would be a server you switched off.
# you never choose a tool from a menu. Your sentence determines whether one matches, the schema determines whether the call is sent, and the limits in FIG.04 determine where it still says no.
$ cat 02-what-this-server-is.md # theseo-mcp 0.1.0
DOC.02

What exactly our server is

theseo-mcp is a remote MCP server. Remote here means that it runs not on your computer but on the internet, and you connect to it via a URL. So you install nothing, download nothing and do not need to have Node or Python on your machine. You paste the endpoint into your client and you are done. The endpoint is https://theseo-mcp.sales-cc8.workers.dev/mcp and that really is the only piece of information you need.

Under the bonnet, it is a Cloudflare Worker, written in TypeScript, that uses the Streamable HTTP transport from the MCP standard

The official MCP TypeScript SDK registers the tools, and the Cloudflare package agents provides the stateless handler around them. Stateless is a deliberate choice here, not a technical detail: a fresh server instance is created for each request, so no shared state can arise between conversations. There is no session memory or user administration. The existing counter measures calls per tool and UTC day. A separately configured outcome measurement may also measure the technical phase, outcome and broad duration band, without input or a user identifier.

That server is deliberately lean

It contains no keys, no passwords and no scanning logic. The actual measurement work takes place on the site Worker for theseo.nl, the same one that powers the free SEO scan and the local findability checker NL on this site. The MCP server is a conduit: it turns your question into a request to a public API endpoint, waits up to thirty seconds for an answer and translates that answer into readable text for your AI. If it takes longer or something goes wrong, you receive an honest Dutch error message instead of a partial result.

Five of the seven tools do not depend on an external network source for their substantive answer

The skill search tool searches a catalogue file supplied with the server, the company-brain template is included as text within the server itself, and the three AI-Index tools read a bundled index of the open data from the Nationale AI-Index. The usage counter operates separately, and a counting failure must not break the tool response. These files are generated from the theseo.nl site repository, so what the server returns matches what was on the site at the time of deployment.

If you visit the server's base URL with an ordinary browser

If you visit the server's base URL with an ordinary browser, meaning the address without /mcp after it, you do not get an error page but a short JSON response containing the name, version, endpoint path and names of the seven tools. It was deliberately built this way: anyone who is curious and pastes the URL should not have to stare at a bare 404.

~/mcp/theseo-mcp/02-route[ok]
Section 02 · The route

From your question to the source

reg. M.002

Three links, that is all there is to it. The client in which you are talking, the MCP server that provides the tools, and the source from which the answer comes. What is recorded where is deliberately stated: the existing counter counts calls, and a separately configured outcome measurement also measures phase, outcome and duration band; for the scan and local check, the site flow stores the address or company context and the result.

FIG.02: Client, server and source[##---] sheet 2/5
01 · CLIENT Where you talk claude.ai and the desktop app
Claude Code in your terminal
Cursor via mcp.json
# any client that follows the standard
→
02 · MCP SERVER theseo-mcp Cloudflare Worker
streamable http at /mcp
stateless, no conversation storage
# technical counts; outcome measurement when configured
# the site Worker in step 03 does record the scan and check
→
03 · SOURCE Where the answer comes from site Worker: /api/scan
site Worker: /api/vcheck
catalogue in the server itself
# Google Places sits behind the site Worker
and the same route back: measured text to your AI, which turns it into a summary that you can verify
~/mcp/theseo-mcp/03-tools[ok]
Section 03 · The tools

Seven tools, input and output

reg. M.003

The input fields below appear literally this way in the code, including their limits. Those limits are not there to obstruct you, but to prevent an AI from accidentally sending half a novel as a company name.

FIG.03: The tools overview[###--] sheet 3/5
tools.json5 read files · 2 CRM operations
seo_scanFree SEO scan (24 points) INPUTurl · text, 1 to 400 characters. The page to be scanned, for example https://example.com OUTPUTAn overall score from 0 to 100 plus four component scores for technical, on-page, content and AI readiness, followed by a line for each point with its status and explanation.
lokale_vindbaarheid_checkLocal findability check (Google Maps) INPUTbedrijfsnaam · text, 1 to 120 characters, as customers search for it. stad · text, 1 to 80 characters, the place where it is based. OUTPUTName and address as known to Google, a score from 0 to 100, the status and weight for each measured criterion, a separate list of what was not measured and why, and specific tips.
skill_zoekenSearch for a free Claude skill INPUTonderwerp · text, 2 to 200 characters, describing what the skill should be about. max · integer from 1 to 20, optional, default 5. OUTPUTFor each match, the name, cluster, an abridged description, the explanation page on theseo.nl and the direct download URL for the zip. At the bottom, the link to the full catalogue.
bedrijfsbrein_templateRetrieve company-brain template INPUTNone. This tool has no input fields at all: you ask for it and you get it. OUTPUTThe explanation page, the download URL and then the full text of both LEESMIJ.md and the fillable bedrijfsbrein.md, ready to be filled in directly within the conversation.
ai_index_brancheAI-Index: who the model mentions in a sector INPUTbranche · text, 2 to 60 characters, for example healthcare or logistics. Part of the name is sufficient. OUTPUTThe number of valid measurements, the refusals, the ten most frequently mentioned companies with the current picture for each question, and how concentrated the mentions are. With a source line.
ai_index_vragenAI-Index: the question list for a sector INPUTbranche · text, 2 to 60 characters. max · integer from 1 to 40, optional, default 10. OUTPUTThe exact customer questions put to the model, plus the fixed prompt, the model and the measurement cycle, so you can reproduce the measurement yourself.
ai_index_bedrijfAI-Index: whether a company name appears in the measurements INPUTbedrijfsnaam · text, 2 to 120 characters, for example Coolblue or bol.com. OUTPUTThe sectors and questions in which the name appears, with counts for each measurement point. If the name does not appear, the tool explicitly states that this is not a judgement.
$ man seo_scan
DOC.03

Tool 1: seo_scan

The first tool scans one URL on 24 points, divided into four categories. Technical covers HTTPS, the HTTP status, server speed, the viewport meta tag, the configured language, robots.txt, the XML sitemap, security headers and email authentication. On-page covers the title, meta description, H1, heading structure, image alt text, canonical and internal links. Content covers the word count, content structure and indexability. And AI readiness covers structured data, the FAQ schema, Open Graph, the Twitter Card and entity clarity.

That final section is why this scan differs from most

AI readiness is about whether a language model can understand and cite your page: does it contain schema markup, is it clear who or what the page describes, and are there fields from which an assistant can compose a clear answer? This is a different kind of work from traditional SEO, and it is rapidly becoming more important as more people ask an assistant their questions instead of using a search bar.

The answer returned by your AI starts with the scanned URL and the time

The answer returned by your AI starts with the scanned URL and the time, followed by the total score, then the four section scores, and then the individual points for each category. Each point has a status: OK if it is configured correctly, INFO if it is for information only, ATTENTION for a warning and CRITICAL if something is genuinely wrong. After that status comes an honest explanation of what is right or wrong, not just a tick. This allows your AI to turn it into a readable summary without having to invent anything itself.

A realistic conversation goes like this

You type: run an SEO scan on my site and tell me what I should tackle first. The AI calls seo_scan with your URL, receives the 24 points and says, for example, that the technical setup is sound, that the meta description is missing from the page you mentioned and that there is no schema markup. You can then simply continue the conversation: ask for the three points with the greatest impact, ask for a suggested better title, or ask for the list to be turned into a task list. That conversation is the benefit. A standalone scanner gives you a list, an AI with a scanner gives you a list plus a plan.

One honest limitation should be mentioned straight away: this scan examines one page and measures only what can be measured automatically

It is not a substitute for an audit. A proper audit examines multiple pages, your competitors, keywords and search intent, and that remains manual work. The tool also says this itself in its answer, with a reference to the scan page for additional context.

Gianluca, who built the MCP server and made it public
image · showing who wrote the server and made the code public, so you can read what each tool does
$ man lokale_vindbaarheid_check
DOC.04
~/mcp · lokale_vindbaarheid_check · example view, fictional company
theseo ~ %
# schematic: the structure of the answer, not an actual measurement
Score achieved from measured points, 0 to 100
[OK] Google Business Profile (20 points)
[OK] Opening hours (10 points)
[NO] Photos (10 points): fewer than ten
[OK] Rating (10 points): at least 4.0
[OK] Number of reviews (10 points): at least 25
[NO] Website linked (10 points)
[OK] Telephone number (5 points)
# not measured, not included: local map pack, responding to reviews
[ok] tips for each point in the answer · without a key: the API error message

Tool 2: lokale_vindbaarheid_check

The second tool looks at how a business is performing in Google Maps. You provide two things: the business name as customers search for it, and the city. Nothing more. The tool looks up the Google Business Profile and then assesses seven criteria that can be measured through the Google Places API: whether the profile exists, whether it includes opening hours, whether there are at least ten photos, whether the rating is at least 4.0, whether there are at least twenty-five reviews, whether the website is linked and whether the telephone number is included. Each criterion has its own weighting, and the answer shows that weighting, so you can immediately see where the greatest gain lies.

status · key missing from the serverThe check is temporarily unavailable. The Google key is not currently on our server, so your AI receives the API error message instead of a score. If you would like us to perform it manually, email sales@theseo.nl.
The most interesting thing about this tool is what it does not do

Two criteria found on many checklists are returned here as not measured, along with the reason. Whether you appear in the local map pack, the three businesses Google displays above the results, can only be determined with a live search. And Google does not make information about whether you respond to reviews available through this route. Both are explicitly returned as not measured and are not included in the denominator of the score. A point that nobody can measure should not lower your score, because then you are measuring your own blind spot instead of the business.

If the business is not found at all

If the business is not found at all, you do not receive a score of zero with an invented analysis, but a short and clear message saying that we cannot find the business on Google, plus the first steps: this usually means that there is no Google Business Profile yet, and creating one is free.

A realistic conversation: check the local visibility of Example Bakery in Bristol

Your assistant calls the tool with the business name and city, receives the measured points and tells you, for example, that the profile exists and the rating is satisfactory, but that there are too few photos and the website is not linked. Then comes the follow-up question that makes this interesting: which two things should I complete this week to increase the score the most? The AI can answer that because it has received the weightings and the tips. Those who prefer to click through themselves can run the same check using the local visibility checker NL.

Another limitation you need to know: this check only works when the corresponding API key is on our server

If it is missing, the tool passes on the honest API error message instead of inventing a result. The boring option, yes, but the only one you can trust.

$ man skill_zoeken # catalogue of 100 skills
DOC.05

Tool 3: skill_zoeken

The third tool searches our catalogue of 149 free Dutch-language Claude skills (the 100 from the library and 49 from articles) by subject. A skill is an instruction file, SKILL.md, that gives an AI a fixed workflow for one task: a SWOT analysis, a quotation, a minutes report, a decision tree. All skills in the skill library are free and require no account. The problem with a library of that size is that you need to know what it contains to get anything out of it. This tool removes that problem: you ask in ordinary language whether there is something for your situation, and your AI searches.

The search tool does this without a network connection, because the catalogue is stored in the server itself

It splits your subject into individual terms and scores each skill in four places: a match in the name carries the most weight, followed by the slug, then the cluster, while a match in the description carries the least weight. This is deliberate, because the descriptions are long and contain many individual keywords. Without that weighting, a skill that happens to mention your word somewhere in its description would rank above the skill that is literally named after it.

For each match, you receive the name and the cluster it belongs to

For each match, you receive the name, the cluster it belongs to, a description truncated at approximately 240 characters, the link to the information page on theseo.nl and the direct download URL for the zip file. By default, you receive five matches, but your AI may request up to twenty. If your search term produces no results, the tool simply says so, suggests a broader term and includes the link to the full catalogue.

A realistic conversation: is there a free skill that helps me with a SWOT analysis?

The AI calls skill_zoeken with that subject, receives the matches and names the most suitable one, including the download link. You can then continue: ask it to download and install the skill, or ask what the difference is between two matches. If you do not yet know how to add a skill to your environment, the page about installing Claude skills NL explains it step by step.

The honest caveat: the catalogue in the server is a snapshot

It is generated from the theseo.nl site repository, so new skills only appear in the search tool after the server has been redeployed. The current list is always available on the site, and that link is included beneath every search result.

~/mcp · skill_zoeken
theseo ~ %
ai-pre-mortem-analist /ai-en-automatisatie/ai-pre-mortem-analist-skill-voor-claude/ · zip
ai-inversie-denker /ai-en-automatisatie/ai-inversie-denker-skill-voor-claude/ · zip
ai-tweede-orde-denker /ai-en-automatisatie/ai-tweede-orde-denker-skill-voor-claude/ · zip
# schematic: for each result, the name, the explanatory page and the zip
[ok] 3 results from 100
$ man bedrijfsbrein_template
DOC.06

Tool 4: bedrijfsbrein_template

The fourth tool has no input fields. You ask for it and receive the complete contents of the free business brain template NL: a fillable context file, bedrijfsbrein.md, that teaches an AI who your company is. Offering, target audience, tone, working method. You fill it in once and then provide it to every AI you work with, so you do not have to explain what you do and for whom again in every conversation.

The response contains three things

The link to the explanatory page, the download URL for the zip, and then the full text of both the readme file and the template itself. The latter is why this tool exists. You do not have to download or open anything: the template simply appears in the conversation, and your AI can fill it in with you straight away.

This is what that sounds like in practice: retrieve the business brain template and fill it in for my painting company

The AI receives the complete template, goes through the sections with you and asks the questions it needs to fill it in. At the end, you have a completed file that you can save and reuse. Feel free to ask it to return it as markdown, so you can paste it directly into a file.

This is also the best example of why MCP is practical

Without the server, you would have to go to the site, download the zip, extract it, open it and paste the contents. With the server, it takes one sentence. It does not save hours, but it removes exactly the kind of friction that causes good intentions to be left undone.

And now the honest account of this template's limitation: it is a text file that you bring with you to every conversation

It does not remember anything by itself, it does not know what you decided yesterday and it does not grow with you. If you want that, you do not need a template but a brain. More on that later, in the section about Jarvis.

$ claude mcp add --transport http theseo-mcp # one line, done
DOC.07

Installing in Claude Code, Claude.ai or another client

There is nothing to install in the traditional sense: you only tell your AI environment where the server is located. How you do that varies by program, but the same URL is used everywhere. Here are the three routes we tested, with the actual commands.

CLAUDE CODE
  1. Run in your terminal: claude mcp add --transport http theseo-mcp https://theseo-mcp.sales-cc8.workers.dev/mcp
  2. Claude Code remembers the server and displays the seven tools in your next session.
  3. Then simply ask for a scan or a skill, you do not need to learn a command.
CLAUDE.AI AND THE DESKTOP APP
  1. Go to Settings and then Connectors.
  2. Choose Add connector and paste the URL https://theseo-mcp.sales-cc8.workers.dev/mcp
  3. Done. For example, ask Claude for an SEO scan of your site.
CURSOR AND OTHER CLIENTS
  1. Open ~/.cursor/mcp.json, or .cursor/mcp.json in your project.
  2. Under mcpServers, add a theseo-mcp key containing url set to the endpoint.
  3. Other clients that follow the standard only need that same URL.
For Cursor, the file will then look like this

If an mcpServers block already exists, place our key alongside the existing ones instead of replacing the block.

# ~/.cursor/mcp.json { "mcpServers": { "theseo-mcp": { "url": "https://theseo-mcp.sales-cc8.workers.dev/mcp" } } }
There is no key, no token and no sign-up step

A deliberate choice: as soon as you need an account to try a free tool, almost no one tries it any more. Because there is no account, you also do not get a personal history. The existing counter measures calls per tool and UTC day. A separately configured outcome measurement may also measure the technical phase, outcome and broad duration band, without input or a user identifier. For seo_scan, the site flow stores the scanned web address and result; for lokale_vindbaarheid_check, it stores the business name, location and result. This recording does not provide a screen where you can look up previous calls next week.

If something does not work, the first check is simple: paste the server's base URL into your browser, meaning the address without /mcp at the end

If you see a short JSON response with the name, version and the seven tool names, the server is running and the problem is in your client. If you see nothing, the problem is on our side.

$ cat 08-why-free.md # MIT
DOC.08

Why this is free and open source

The most honest reason first: we do not make our money from scans. We make it from the work that follows and from our software. A free tool that genuinely measures something is the cheapest way for us to show how we think. Anyone who runs the scan and reads the response can see within a minute whether they like our way of looking at things. A better introduction than any brochure, as far as we are concerned.

The second reason is that a closed measurement tool has hardly any value

If you cannot check what is being measured, you have to take our word for it that a score is correct. That is why everything is open: the server's complete source code is available at github.com/GianlucaNaarden/theseo-mcp under the MIT licence. You can read exactly which fields a tool expects, what happens to your input and how the response is constructed. If you think something is nonsensical, you can open an issue. If you want to run or modify it yourself, you are free to do so: the MIT licence permits that, including commercially.

Running it yourself is not just theoretical

You need Node 20 or newer, then you run the tests, start it locally and deploy it to your own Cloudflare account. The tests are a story in their own right, by the way: they use the official MCP client to communicate with the real handler, in the same process and without a network. That means they test what actually happens instead of a reconstructed version of it.

$ npm install $ npm test # smoke test plus tests for the skill finder $ npm run dev # locally at http://localhost:8787/mcp $ npm run deploy # to your own Cloudflare account
The third reason is that it is simply fun to make something new

As far as we have been able to determine, this is the first MCP server from a Dutch marketing agency. That caveat is necessary, because we cannot search every register and repository in the world, and someone else may already appear there tomorrow. What we do know for certain is that the standard is young, there are still few Dutch-language tools, and we thought it would be a shame to wait until someone else built it.

What we gain from it is no secret

Anyone who finds this tool useful will naturally come to us at some point for the work we do charge for: SEO, Google Ads, AI and automation and our own software NL. That is the entire arrangement, and it is stated here in black and white so you can verify it.

~/mcp/theseo-mcp/04-limits[ok]
Section 04 · The limitations

What this server does not do

reg. M.004

Any tool that aims to be honest must know its own limits. Below are the requests you might make and the answer you actually receive, including the two criteria deliberately excluded from the score because they cannot be measured.

FIG.04: limits.log[####-] sheet 4/5
limits.logthe limits of the server, set out in writing
perform a complete audit of my entire websiteLIMITEDThe scan examines one page for 24 automatically measurable points. It provides a useful first impression, not a complete audit. Multiple pages, competition, keywords and search intent still require manual work.
do I appear in Google's local map pack?NOT MEASUREDThat requires a live search query. This is not included in this free check. The criterion is returned as not measured, together with the reason, and does not count towards the score.
do I respond to enough of my reviews?NOT MEASUREDGoogle does not make that available through this route. The tool says so as well, with the recommendation that you should do it: responding is one of the strongest activity signals.
save my scan results for next weekNOT POSSIBLEThis server does not provide personal memory. There is no user database, account or session in which you can retrieve previous results. The existing counter measures calls per tool and UTC day. A separately configured outcome measurement may also measure the technical phase, outcome and broad duration band, without input or a user identifier. For the SEO scan and local check, our site records the scanned address or business context and result in our own database, just as it does for the tool on theseo.nl. That is our administrative record, not a history you can request.
create an account for meNOT REQUIREDThere is no account and no key. You paste in the endpoint and it works. So there is no account password or access token here for you to remember or for someone to leak from such an account.
is last week's skill already included?SNAPSHOTThe catalogue is supplied with the server. New skills appear in the search tool after a new deployment. The current list is always available in the skill library, and that link is included with every result.
the visibility check returns an errorHONEST ERRORThe check only works when the API key is present on our server. If it is missing, the tool passes on the API's error message verbatim instead of inventing a result. Dull, but verifiable.
~/mcp/theseo-mcp/05-upgrade-path[ok]
Section 05 · From tool to employee

Run it yourself or have it run for you

reg. M.005

To be completely honest: this server is the free do-it-yourself version of work we also provide as a service. It remains complete and free of hidden catches, but understand what a tool is. It responds when you ask, and every new session starts empty. The tool is not the engine and not the memory. You ask, you provide the context, you check. Anyone who wants this handled differently has two next steps: hand over the engine, or arrange the memory.

$ cat from-tool-to-employee.mdthree tiers, same work
upgrade-path.shfree · employee · brain
TIER 1 · FREE
where you are now
The MCP server: you are the engine You connect the server to Claude, Claude Code or Cursor and request a scan, a check or a skill yourself. It costs nothing, works today, and everything on this page will remain available just as it is. The limit is your own time: it happens only when you ask. $ claude mcp add --transport http theseo-mcp · 0 euro · you ask, you check
↓
TIER 2 · SERVICE
have it prepared
The AI employee: it is ready without you having to ask Exactly this kind of work, but as a service: the Reporting Employee runs the measurements and prepares the weekly and monthly reports without you having to sit down at a terminal. You retain control, because output remains a draft until a person approves it. We provide this through Mansotti, the business whose trading name is TheSEO, which builds a Quotation Employee and a Sales Employee alongside the Reporting Employee. Read what an AI employee is and does. from 950 euro per month · a person always gives approval
↓
TIER 3 · BRAIN
everything from one source
Jarvis: all your AIs work from the same business knowledge A tool teaches your AI what it can retrieve, while the brain is where the memory lives. If you want all your AIs to work from the same business knowledge, that is Jarvis, the organisational brain. It connects ChatGPT, Claude, Codex and your people to the same projects, core knowledge and decisions, so your next AI session does not start over. Jarvis selects context, delegates work and stores only what has been approved: client knowledge remains isolated and every step leaves a verifiable trail. This server benefits from that too, because you no longer have to repeat which sites and which businesses you are monitoring in every conversation. Read what that delivers in concrete terms, from the plans to your first week, on Jarvis itself. entry-level Brain Start: 9 euro per month incl. VAT · pay for your brain, not per AI request
# no advertising trick: tier 1 remains free and complete. The following tiers are for those who want to hand over this work.
~/mcp/theseo-mcp/06-jarvis[ok]
Section 06 · The brain

What Jarvis delivers in concrete terms

reg. M.006

Tier 3 deserves more than one paragraph, particularly on this page. Anyone who connects an MCP server understands the distinction between a model that guesses and one that can retrieve something. But there is a gap behind it that you only notice after a few weeks. By design, the MCP server does not store conversations: the technical measurement contains no conversation content or user identifier. For the two forwarding tools, the site flow does store the address or business context and result, but not what you decided afterwards. That decision evaporates as soon as the conversation closes. Jarvis is the layer that retains exactly that.

FIG.05: The tool provides the fact, the brain retains the decision[#####] sheet 5/5
work-journal.mdexample layout, no client data
what the tool returnedwhat was decided afterwards
seo_scan · week 1H1 is missing from the service page, meta description too longnot recorded as a decision, only discussed
lokale_vindbaarheid_check · week 3opening hours are missing for two branchesnot recorded as a decision, only discussed
seo_scan · week 5H1 is missing from the service page, meta description too longthe same finding, the same discussion, no idea whether anything had already been decided about it
$ jarvis on · the same tools, the same measurements, but now the right-hand column remains
seo_scan · week 1H1 is missing from the service page, meta description too long[decision]H1 will be the service name plus the location; only rewrite the meta after that page has been reorganised
lokale_vindbaarheid_check · week 3opening hours are missing for two branches[task]requested from the branch manager, draft ready, awaiting human approval
seo_scan · week 5H1 is missing from the service page, meta description too long[recall]the decision from week 1 is included: the meta is deliberately on hold, so this is not a new finding but an outstanding one
# the MCP server changes nothing about its work here: in week 5 it measures exactly the same thing as in week 1, as it should. The difference is in the adjacent column. The complete trail from recording to approval is available at jarvis/werking.
mcp.plus Last month's measurement is still there The server is stateless, and that is a security choice we do not want to abandon. The brain therefore sits alongside it, not inside it: that is where what you did with a measurement is recorded, so the finding from week 5 is recognised as the outstanding one from week 1.
tools.shared The same tools in every AI you use You paste the address separately into Claude, Cursor, Claude Code and ChatGPT. What you learn and decide there then remains separate. With a brain behind them they read from the same documents, and one AI knows the decision you recorded with another.
scan.schedule A measurement that comes back on time Calling it manually works perfectly well as long as you remember. With a recurring task in the team space the next measurement is waiting; your AI sees it in the task list and runs it then. Jarvis itself runs nothing: the AI you connect does the work, the task makes sure it is not forgotten.
trace.back Trace which call supported which decision The log in the team space keeps what was done and decided, and every document in your brain keeps its earlier versions. Without that trail, six months later you can no longer determine what a choice was based on.
human.approval Saved when you ask Your AI saves a document in your brain when you ask it to, and a change is made on top of a known version, so two AIs do not silently overwrite each other. Jarvis itself sends no email, publishes nothing and pays nothing. You can see every saved version, and a deleted document can be restored.
brain.isolated Your own, separate brain Your brain belongs to your own account, and an AI only sees what you linked to that connection. Several brains, for example one per client, come with Team and Business, which cannot be ordered yet.
# THE HONEST EVIDENCE · NOT A DEMO

We ourselves have already been running on this system for months. Every agent session, every task and every decision is logged in it and can be reviewed. As a result, a new session does not start with a blank slate: it first retrieves the recorded decisions, ongoing projects and latest changes, then continues where the previous session stopped. This server itself emerged from it, including the decision to keep it stateless. We are not describing a promise here, but the way we ourselves work every day.

$ cat pricing.mdpay for your brain, not per AI request
Brain Start €9 /month incl. VAT 1 user · 1 own brain · 2 AI connections · 7 days of history
Brain Solo €29 /month incl. VAT 1 user · 1 own brain · 3 AI connections · 30 days of history
Brain Team €99 /month incl. VAT cannot be ordered yet
Brain Business €249 /month incl. VAT cannot be ordered yet

View the four plans at jarvis/prijzen NL. Order Start and Solo online at Start with Jarvis NL; Team and Business cannot be ordered yet. We discuss bespoke business solutions first.

~/mcp/theseo-mcp/07-continue[ok]
Section 07 · Keep exploring

What this server relies on

reg. M.007

Every tool on the server has a page on this site where the same work can be done manually. That is not duplication but a choice: some people want a button, others want a conversation. The entire range is brought together below.

The sources behind the tools

same measurement, different route

These are the scanners and files that the seven tools rely on. They all work without AI too.

Free SEO scanThe same 24 points as seo_scan, but with a button and an explanation for each point on the screen.TOOL Local visibility checkerThe same check as lokale_vindbaarheid_check, including what is deliberately not measured.TOOL NL The skill libraryThe 100 free skills of the library, which skill_zoeken searches together with 49 from articles, always up to date and sorted by topic.HUB Company brain templateThe file returned by bedrijfsbrein_template, with an explanation and download.DOC NL

Get started with AI yourself

from tools to working method

Anyone who enjoys setting up this kind of thing themselves will find the next step here.

Installing Claude skillsHow to add a SKILL.md to your environment, step by step, including for Codex and Cursor.DOC NL AI and automationThe service behind it: from individual tools to working automation in your company.SRV AI trainingFor teams that want to learn how to set up and manage this kind of work themselves.SRV Knowledge baseSearchable articles about SEO, AI and online visibility.DOC

Our own software

what else we build

This server is not the only thing we make ourselves. The full overview is on the software page.

Software by TheSEOAll the tools we build and operate ourselves, with an explanation of each one.INS NL JarvisThe organisational brain that connects your AI tools to the same company knowledge.INS The AI employeeThis work as a service, with final human review before anything is published.SRV AI indexMeasures which companies AI mentions by sector, based on open data.DOC NL

The work behind it

how we make our money

The scan points something out to you, and then the real work begins. We can simply do that for you too.

SEOBeing found consistently, in Google and in AI answers.SRV Google AdsAdvertising that delivers returns, with measurements you can verify yourself.SRV About TheSEOWho we are, how we work and why we give away this kind of thing.DOC ContactQuestions about the server or ideas for an eighth tool: let us know.DOC
$ claude # interactive session · eight questions, eight answers
DOC.09 · FAQ

Frequently asked questions

What exactly is MCP?

MCP stands for Model Context Protocol. It is an open standard that defines how an AI assistant may use external tools. On its own, an AI model can only produce text based on what it has read in the past. MCP allows it to request something from a server that can measure, search or retrieve information. The server explains which tools are available, what input they expect and what they return, and the assistant decides when to call one.

How much does this MCP server cost?

Nothing for the thirteen tools without an account on mcp.theseo.nl and for the seven tools of the open base: no account, no email address and no key. The code of the open base is on GitHub under the MIT licence. It only costs money if you choose Jarvis Start (9 euros a month) or Solo (29 euros a month), including VAT, or the premium skills; a measurement on request is deducted from your account credit. No sales email will follow.

Which programs does the server work in?

In any client that supports remote MCP servers via Streamable HTTP. We document Claude on claude.ai and in the desktop app, Claude Code, and Cursor via the mcp.json file. Other clients that follow the standard only need the endpoint URL, as that is the only thing you need to enter.

Does the server store my data?

For each valid tool call, the existing usage counter receives the tool name, UTC day and the number 1. When the separate outcome measurement is configured, it also receives a fixed schema version, whether the call started or finished, its status as pending, successful or failed, and a broad duration band. The bands are under 1 second, 1 to under 5 seconds, 5 to under 30 seconds, 30 seconds or more, or not measured. These measurement points contain no tool input, response text, error text, IP address, headers, account data, conversation content or user identifier. We do not link them to people or individual conversations. Cloudflare Analytics Engine automatically adds a technical timestamp and, according to its documentation, retains the points for three months. We use the measurement for call counts and technical reliability, not for unique users or commercial results. Missing data and sampling are identified as limitations. A configured binding alone does not prove that points have been stored. This is separate from the existing site flow: seo_scan stores the scanned web address and result; lokale_vindbaarheid_check stores the business name, location and result. This is the same recording as for the site forms on theseo.nl and may create a lead in our CRM. The other five tools read bundled files. Technical measurement does not provide permission to store personal context or conversations in the CRM.

Does the SEO scan replace a proper audit?

No, and the tool itself says so in its response. The scan checks a page against 24 automatically measurable points and gives it an honest score. A full audit examines multiple pages, competitors, keywords and search intent, and that remains manual work. Think of the scan as a useful first impression that tells you where to look.

Why are some criteria returned as not measured?

Because we would rather say nothing than make something up. The local visibility check measures what is available through the Google Places API. Whether you appear in the local map pack requires a live search, and Google does not disclose through this route whether you respond to reviews. Those two are returned as not measured, together with the reason, and they do not count towards the denominator of the score. This means an unmeasurable point can never lower your score.

Is the newest skill included immediately?

Not automatically. The skill catalogue on the server is a snapshot generated from the theseo.nl site repository. New skills appear in the search tool as soon as the server is redeployed. The complete and always up-to-date list is on the skills page of theseo.nl, and that link is also included at the bottom of every search result.

Can I run or modify the server myself?

Yes. The code is available on GitHub under the MIT licence. You need Node 20 or newer, after which you can run it locally with npm run dev and deploy it to your own Cloudflare account with npm run deploy. The server itself contains no keys: the scan and check logic runs on the theseo.nl site Worker, and this server only calls its public endpoints.

DOC.10 · Next step

Connecting takes a minute

Paste the endpoint into your client, ask it to scan your own site and see what comes back. If you like the approach, you will immediately know how we work. If you do not, it has cost you one minute and there is no account in your name. That is the entire trade-off.

If a tool that belongs here is missing, or you encounter something that is not right, let us know. The code is open, so creating an issue on GitHub is the quickest route. If you would rather discuss what else AI could do for your company, from individual tools to an organisational brain, we can simply do that in a conversation.

100 skillsFree template NLJarvisAI Index NLFree SEO scanVisibility check NLSoftware NLInstall skills NLAI employeeAI and automationAI policyDisclaimer NLComplaints NLKnowledge baseContact
Section 08 · Next stepreachable 24/7 · sales@theseo.nl
Book a call ↗ NL
$ whoami
Written by GianlucaFounder of TheSEO. Has worked on visibility for Dutch companies since 2017, in Google and in AI answers. More about the institute.

Back to the short version