The data problem for AI in African agriculture is not a lack of data.
It is a structural absence of structured data — the kind that AI agents can query, validate, and reason with.
The raw material exists. It has existed for over a century. It sits in the public domain, digitized by libraries and archives, waiting to be structured.
The British East Africa Protectorate agricultural reports from 1910-1922 document, county by county, which crops grew where, what yields looked like in drought years versus good years, which soil types supported what cultivation, and which traditional practices survived contact with colonial agricultural extension.
The Kew Gardens Economic Botany Series from 1906-1920 contains botanical surveys of East African crops — sorghum, millet, cassava, cowpea, moringa — with documented water requirements, storage characteristics, and soil tolerances.
The East Africa Meteorological Department records from 1900-1940 show rainfall patterns before the satellite era: where the long rains reliably arrive, where they don't, where the variability is structurally high.
These documents are free. They are in the public domain. They are not behind paywalls. They are in libraries.
The problem: they are in the format of 1910.
Scanned PDFs. Handwritten tables. Narrative reports in colonial English. Measurement systems that predated standardization.
No AI agent built in 2025 can query them the way it needs to. "What is the planting calendar for sorghum in semi-arid Kenya?" requires that someone has extracted, structured, validated, and made available the answer in a format that agents can use.
That structured layer is what's missing. Not the underlying knowledge.
The MCP servers in the East Africa coordination stack (wapimaji-mcp, kilimo-mcp, bima-mcp) need grounded knowledge to reason correctly.
When a farmer asks an AI agent about optimal planting dates for a semi-arid county, the agent needs historical knowledge of rainfall patterns and crop tolerances — not just a general LLM approximation.
That knowledge exists in the 1910 Kew surveys. The work is structuring it so the MCP can use it.
Six public domain datasets now available on HuggingFace:
All public domain. All structured for AI agent use. All connected to the coordination infrastructure stack.
The knowledge infrastructure problem in Africa is not a lack of knowledge. It is a lack of structured access to knowledge that already exists.
Public domain works published before 1928 are the closest thing to free raw material in the information economy. The opportunity is not to extract value from them. It is to structure them so they can flow. These datasets are available at huggingface.co/gmahia. Related coordination tools at pypi.org/user/gmahia.