Query it live

What this is

Everything on this site is also served over MCP, which is the standard way an AI assistant is given something to read that its training did not cover. Connect it once and the directory becomes a thing your assistant can consult mid-conversation, the way it would consult a web page, except that it arrives structured and it arrives with its limits attached.

It is the same data as the download, moving in the other direction. The files come to you and sit where you put them; the connection stays here and answers questions. Nothing goes the other way: the endpoint is read-only, it takes no account, it stores nothing, and it never sees who is asking. There is no analytics behind it and no log worth the name, which is a privacy property rather than a boast about restraint.

It is not a chatbot and it does not answer questions itself. It hands your assistant rows. What the assistant then does with them is between the two of you, and the section below on what this will not do is the part worth reading before you trust an answer.

Connect it

The URL

https://games.thisisdelightful.com/mcp — that is the whole configuration. There is no key to request, no header to set and nothing to fill in beside it.

In Claude

Open Settings, find Connectors, and choose to add a custom connector. Paste the URL, give it a name you will recognise, and save. Claude will show you the six tools it found; leave them all on. The exact wording of those menus moves between releases and platforms, so hunt for the word Connectors rather than for the sentence written here.

In anything else

Any MCP client takes the same URL. The transport is Streamable HTTP and the server is stateless, so there is no session to establish and no order you have to call things in. If your client asks which transport to use, it is HTTP rather than stdio, and it is not SSE.

Checking it worked

Ask your assistant to list the tools it now has from the connector. Six names should come back: search_sources, get_source, get_market, list_trends, get_landscape, search_techniques. If none do, the connector is not attached, and anything the assistant then tells you about the games industry is coming from memory rather than from here.

What to ask it

Once it is connected you do not have to name the tools, and you should not try to. Ask the research question and let the assistant pick. What is worth doing instead is telling it, once, what the directory is for — that it is a map of where to look rather than a store of facts, that nothing in it is ranked, and that it should quote the rows rather than paraphrase them.

The four below do that. Take one, change the market or the question, and go. There are longer prompts written for the downloaded files on Get the data, and everything they say about what a model must not do applies here too.

• Set it up, once per conversation

  Copy

  You have a connector to Delightful's Game Research Starter Pack, a directory of where to research the video games industry. Use it rather than your own memory for anything about games industry sources, markets, trends, platforms or genres. Three rules for the whole of this conversation. First, it is a map of where to look and not a store of facts: every source in it was opened and read before it was described, but nothing the sources themselves claim has been checked, so treat a description as a description and never as a finding. Second, nothing in it is ranked or scored and row order means nothing, so do not present results as a top ten or imply that the first row is the best. Third, when you use it, name the rows you used and give me the link the row carries. If the directory does not cover what I asked, say so plainly rather than filling the gap from what you already know. Start by telling me which tools the connector gave you and, in one line each, what you would use them for.

• Build a reading list for one market

  Copy

  Using the games research connector, find everything covering Japan. Group what comes back by the group column, and inside each group put the sources I would be reading in Japanese first. Give me a table of the name, what it lets me find out, and the link. Then tell me which languages I am going to need, and flag any group holding fewer than five sources, because that is thin coverage in this directory rather than a fact about Japan. Do not add sources from your own knowledge, and do not tell me which of them is best.

• Find who publishes a number

  Copy

  I need to state how much was spent on mobile games in Germany last year, and I need something citable. Using the games research connector, list every source that could plausibly publish that figure, starting with the data platforms and the trade institutions. For each, tell me what the directory records about who owns it, because several of these are owned by the companies they report on. Cost is not recorded on every row: where the connector returns no cost, say the directory does not record it rather than reading that as free. Then order them by how likely I am to get the number without a subscription, and say where you are ordering on a gap. Do not give me the figure itself. I am going to go and read it.

• Brief me on a market before a call

  Copy

  Using the games research connector, pull the full profile for Korea and the trends that land there. Write me a one-page brief: what the market is known for, how its press is structured, where its data lives, and which two or three trends I should expect to come up. Quote the sourced claims as the profile states them and keep the link on each one. The profile carries a player-population figure with a definition of who was counted and a market-size figure published on a particular basis: give me both, give me the definition alongside, and do not compare either against another market, because these figures are not built to be comparable and the connector will have told you so.

The six tools

search_sources

The directory itself, and the tool almost every question ends up in. Free text over the name, the description and the owner, plus filters for market, group, medium, language and cost. Answers where would I look for this. Filtering on cost narrows to the rows where cost was asked at all, which is the data platforms and not the press outlets or the individual people.

get_source

One source in full, by a permanent slug that search_sources hands back. The slug names the same source across releases, so it is the thing to cite if you are building something that has to still work next year.

get_market

A full profile of one of the ten profiled markets: the sourced narrative with a link on every claim, the market-size figure with the basis it was published on, the player-population figure with the definition of who was counted, and where that market's data lives. Sources carry tags for more markets than this, and those tags reach search_sources rather than a profile.

list_trends

What is currently moving the industry, filterable by market, by stage and by category, each with what it is, why it matters and dated further reading. This is the fastest-decaying thing here and every row is dated for that reason.

get_landscape

The platform map or the genre map. Platforms with who owns each one; genres with definitions, subgenres, example games and the terms actually used in Chinese, Korean and Japanese. Worth reading before comparing two datasets, because the industry has no shared taxonomy and this is where the disagreements are written down.

search_techniques

Ways of finding things, each with a query you can paste and an explanation of what it surfaces. Includes local-language strategies for researching Japan, Korea and China without reading the language.

What every answer carries

Every response comes back wrapped in the same envelope, and the wrapping is the point. It carries the date the directory was last compiled, the CC BY 4.0 attribution, a link to the page on this site the rows came from, and the standing caveats. So a figure lifted out of it arrives with the reason it cannot be compared across markets still attached, rather than losing that on the way into a slide.

The caveats are three, and they are the same three every time: this directory does not report, analyse, rank or score anything; market-size and player-population figures are not comparable across markets, because each is published on a different basis, in a different year, against a different definition of a player; and the industry has no shared genre taxonomy, so two datasets that look like they describe the same thing often do not.

Some answers add a note about themselves. Filter on a market that tags sources but has no profile behind it and the answer says so. Filter on a field that is only recorded on some rows and the answer says how many rows it could not speak for. That is deliberate: a short confident list is the failure mode of a filtered directory, and a model has no way to notice it unless the data tells it.

What it will not do

It will not tell you what is true about the games industry. It tells you who publishes on it. Every source here was opened and read before it was described, and what the source then claims is not checked by anyone. A number you reach through this connection carries exactly the authority of whoever published it, and passing it through an assistant does not upgrade that by a single point.

It will not rank anything. There is no score in the data, so there is nothing for a tool to sort on, and an assistant that hands you a best five has invented the ordering. If you get one, ask where the ranking came from and watch it come apart.

It will not fill a gap. Where a field was never collected for a kind of row, the field is absent rather than empty, so an assistant cannot read a blank as a value. Where the directory is thin on a market, it is thin, and the honest answer is that this data has little to say rather than a confident short list.

It will not stay fresh on its own. The directory is compiled in passes and the date it was compiled is on every answer and at the top of this page. The trends decay quickest. If you are reading this a long way after that date, treat the trends as a record of what was being argued about then.

For anyone building on it

The endpoint is the address given under Connect it above. It is Streamable HTTP and stateless: POST JSON-RPC 2.0 to it, and note that a GET returns 405, because nothing here pushes and there is no event stream to open. That is also why the address is written as text on this page rather than as a link — following it in a browser would only produce the 405. CORS is open, so a browser can call it directly. Nothing is rate limited today, and if that changes it will be because somebody made it necessary.

The tool definitions are not written out on this page, and that is on purpose: a hand-copied schema is a schema that goes quietly out of date. POST a tools/list and read the current ones, including the enumerated values each filter accepts, which come out of the directory's own vocabulary rather than being typed anywhere.

Read the catalogue first if you are deciding what to build. It says what tables exist, how many rows each holds, how many rows carry each field, and which market codes have a profile behind them — which is the fastest way to find out whether the filter you were planning to lean on is populated enough to lean on.

• Server discovery document, naming the endpoint and the version
• The catalogue: what exists, how many rows, and how well each field is filled
• llms.txt, the same directory written for a model reading the site itself

The other ways in