Ask AI

You can now explore the Islam West Africa Collection (IWAC) by asking questions in plain language, using the AI assistant you already work with. Two open components make this possible: a read-only MCP server that gives a compatible assistant structured access to the collection, and a research skill that teaches it how to search well. This page explains the approach and shows how to connect Claude or ChatGPT.

Why keyword search falls short

Keyword search is fast and familiar, but it matches strings rather than ideas. A query for "hajj" might not surface articles on "pilgrimmage to Mecca" unless those exact words appear, and the same string "hadj" returns both the pilgrimage and the many men who carry "El Hadj" as an honorific. Historical sources compound the problem with spelling variants, language change, and concepts that do not map onto present-day terms. OCR noise from old newspaper scans degrades the text further. As Huistra and Mellink (2016) observe, mass digitisation pushes researchers to approach sources "through specific words", which leaves results at the mercy of the search tool. Keyword search also assumes you already know how the collection is organised, which terms it uses and which countries it covers well. Most users do not.

Why not just ask a chatbot?

People increasingly put their questions straight to a commercial AI chatbot instead of a search box. Ask one about Islam in West Africa, though, and the answer is usually generic and sometimes stereotyped: the region is thinly represented in the data these models were trained on, so they paint in broad strokes and, at worst, invent their sources. The IWAC is open access, and over the past year AI crawlers, many of them from the United States and China, have repeatedly harvested the site. Yet for all that traffic, the chatbots barely draw on the collection and rarely cite it. They can scrape the IWAC; they cannot use it well.

A different way in: a server and a skill

Rather than a single chatbot, the IWAC opens to AI through two layers that work with the assistant of your choice.

The MCP server — the plumbing

The Model Context Protocol (MCP) is an open standard, released in November 2024, that connects an AI assistant to an outside source of data through a defined set of queries called tools. If you have ever let an assistant read your email or calendar, you have used an MCP server. The IWAC server offers more than twenty read-only tools: search the newspaper articles, read one in full, look up a person or organisation, compare coverage across countries, consult the sentiment analysis already computed for each article. The principle is the same as IIIF for digital images: a shared interface that lets others use the collection without taking custody of it.

The research skill — the method

A SKILL.md is a set of plain-language instructions that teaches AI agents how to use those tools well. The IWAC skill encodes a five-phase research method drawn from archival practice: establish what the collection holds before searching; search systematically in French, with transliteration variants for Arabic-Islamic terms (Tabaski for Eid al-Adha, Maouloud for Mawlid); read the strongest sources in full; cross-check across articles, publications and references; and report findings with their sources and a confidence grade. The skill also states the collection's gaps in every answer, from its francophone tilt to the thin coverage of Niger and the absence of Nigerian press, so the assistant never passes a partial corpus off as the whole picture.

Every result links back to the original record on islam.zmo.de, so you can check any claim against its source. The server generates nothing when you ask a question; it only looks up what the collection already holds. The assistant works from the archive itself rather than from its training data, so its answers stay anchored to sources you can verify.

What you can ask

Once connected, you can put questions such as these to the collection:

"How did the Béninois press cover Tabaski (Aïd al-Adha)?"
"Compare newspaper coverage of Islam across Côte d'Ivoire, Burkina Faso and Benin."
"Find academic references on Izala in Niger, in French and English."
"Summarise what the collection holds on the debates oversecularims in the 1990s."

You ask in plain language, and the assistant works out which tools to call and how to combine the filters. You can ask in French or English. Most newspaper sources are in French, and the assistant searches them in French whatever language you use.

How to connect

There are two ways to connect, depending on the assistant you use.

Claude Desktop (runs on your computer)

Recommended for most users. The install is one click, with no command line, and it includes the full toolset plus the research skill.

Open the releases page and download the bundle for your operating system: iwac-mcp-server-windows.mcpb for Windows, or iwac-mcp-server-macos.mcpb for macOS. (Claude Desktop has no Linux version, so there is no Linux bundle.)
Double-click the file. Claude Desktop opens an install dialog; review the details and click Install. You can also drag the file onto Settings → Extensions.
The first time you ask a question, the server downloads about 250 MB of data from Hugging Face onto your computer. After that, searches are fast and fully local.

Add the research skill (strongly recommended). Download iwac-mcp-skill.zip from the same release. In Claude Desktop, open Customize → Skills → + → Create skill → Upload a skill and select the zip. With the skill installed, you simply ask your question and Claude plans the searches for you.

Optional — search by meaning. Twenty tools work out of the box. Two more find articles and publications by meaning (semantinc embeddings) rather than by keyword, so they surface a piece about an idea even when it never uses your exact words. They are off by default and need a free Google AI Studio key, which you paste into the extension's settings. Most users do not need this.

ChatGPT

The server is also hosted online, so you can add it to ChatGPT as a custom connector. You need a paid plan (Plus, Pro, Business, Enterprise or Edu) and ChatGPT on the web at chatgpt.com. The connector is read-only and needs no login and no API key.

Turn on Developer mode under Settings → Apps & Connectors → Advanced settings. Depending on your account, this section may be labelled "Apps" or "Connectors".
On Apps & Connectors, click Create, then enter:
- Name: IWAC
- MCP Server URL: https://islam.zmo.de/mcp/
- Authentication: No authentication
Save. ChatGPT connects to the server and lists its tools.
In a chat, open the + menu → Developer mode and tick IWAC, then ask your question. Naming the connector in your prompt helps: for example, "Using the IWAC connector, find press coverage of…". ChatGPT's deep research mode can use the same connector to find and read sources and cite IWAC links in its report.

ChatGPT cannot load the research skill. The server carries the most useful guidance, including French transliterations and the coverage caveats.

What it does, and what it does not

This approach widens access. You can ask in natural language, follow a question wherever it leads, and reach material that keyword search would miss, with every claim traceable to its source.

It does not replace careful research. Grounding the assistant in the collection reshapes the risk of error rather than removing it: the assistant can still misread a source, treat distinct terms as interchangeable when they are not (intégrisme, fondamentalisme and salafisme carry real differences), or imply that the archive speaks for itself. The answers are only as good as the metadata beneath them, and the collection remains partial, shaped by what was digitised, in which language and from which country. It supports exploration and distant reading, and it complements close reading and your own judgement.

What's next

A conversational assistant is planned for the IWAC website itself, running an open-source model against these same tools, so that consulting the collection will need neither a subscription nor an account with a commercial AI tool.