Three protocols landed in 2024 that change what AI agents can do in production: MCP (Model Context Protocol), A2A (Agent-to-Agent), and Google ADK (Agent Development Kit).
The question for Africa's coordination infrastructure is not whether these protocols are technically interesting. The question is: what does the absence of domain-specific implementations mean for the coordination problems — health, water, agriculture, land, education — that remain unsolved at scale?
The structural absence
Most MCP implementations connect to platforms that already exist in rich countries: Slack, Google Calendar, GitHub, Stripe, Salesforce. The ecosystems those tools coordinate are already coordinated. The problem is solved.
The structural absence is in the domains where coordination is genuinely broken:
- A community health worker in Turkana County manages 1,200 households with a paper register. No MCP server for CHW coordination.
- A smallholder farmer in Embu has no way to know today's market price at the Meru collection point before deciding to sell. No MCP server for farm-gate price information.
- Drought early warning data exists (NASA NDVI, NOAA CHIRPS) but there is no structured AI-accessible interface to it. No MCP server for drought intelligence.
These are not technical problems. They are data structuring and interface design problems. The protocols exist. The domain implementations don't.
What the stack looks like
The East Africa coordination stack now includes 31 MCP servers:
- Payments: M-Pesa API via Daraja (mpesa-mcp)
- Water/drought: Satellite data integration (wapimaji-mcp)
- Agriculture: Market coordination, crop advisory (kilimo-mcp, soko-mcp)
- Parametric insurance: Satellite-triggered coverage (bima-mcp)
- Health: County health service coordination (afya-mcp, county-mcp)
- Land: Registry coordination (ardhi-mcp)
- Tax compliance: Kenya Revenue Authority (kra-mcp)
- Capital markets: NSE data access (faida-mcp)
- Labor rights: Worker compliance (haki-ya-kazi-mcp)
- Language: Swahili NLP infrastructure (swahili-civic-nlp, tafsiri-mcp)
The A2A pattern enables agents built on these servers to coordinate without bespoke integration code. The ADK pattern provides the framework for agents that can plan multi-step tasks — necessary for coordination workflows spanning multiple API calls and decisions.
The architecture gap this addresses
The composable model future assumes domain knowledge is accessible to AI agents through structured interfaces. For most African coordination domains, it is not.
The problem is not a model problem. GPT-4, Claude, and Gemini are accessible. The problem is that there is no structured interface between those models and the data that would make them useful for a CHW in Turkana, a farmer in Embu, or a water officer in Wajir.
MCP, A2A, and ADK are the protocol layer. The East Africa implementations are the domain layer. The work is connecting them.
All servers available at pypi.org/user/gmahia and indexed at Glama.
Top comments (0)