The East Africa Coordination Mafia: Why the 23 MCPs Are a Platform, Not Just Tools
"The real asset isn't the company. It's the apprenticeship."
There's a well-documented pattern in tech: elite companies produce second-generation founders who don't replicate the original company — they export its epistemology. SpaceX alumni are building transfer vehicles, satellite buses, and reusable rockets. PayPal alumni built LinkedIn, YouTube, and Yelp — not more payment processors. OpenAI alumni founded Anthropic, Thinking Machines Lab, and Eureka Labs.
Peter Thiel's observation: people copy organizational culture more than technology. What PayPal exported wasn't code. It exported speed, ambition, willingness to attack monopolies.
What SpaceX exported wasn't rocket blueprints. It exported first-principles thinking, vertical integration, and manufacturing discipline.
I've been thinking about this pattern in the context of what I'm building in East Africa.
The Coordination Infrastructure Layer
The 23 MCP servers I've shipped aren't really 23 separate products. They're a single answer to one question:
Why can't an AI agent in Nairobi do what AI agents in London, New York, or Seoul can do?
Not because of compute. Not because of data. Because the institutional knowledge that makes those agents useful — tax codes, labor rights, crop calendars, insurance frameworks, water access data — lives in silos, in PDFs, in government portals. It's not composable. It's not API-accessible. It's not AI-native.
kra-mcp is the TurboTax model for Kenya. faida-mcp is the Fidelity model. kilimo-mcp is Climate Corp. haki-ya-kazi-mcp is ACAS. afya-ya-akili-mcp is BetterHelp.
Each one encodes the bottleneck insight: institutional coordination knowledge that should be universally accessible isn't. MCP makes it composable.
What This Stack Actually Exports
If other developers build on this stack — and they will — they won't be building more MCP servers. They'll be exporting:
Coordination epistemology: The understanding that Africa's institutional gaps are primarily information asymmetry problems, not capital problems. This is a paradigm shift. Once you see it, you can't unsee it.
Domain-first AI methodology: Build the domain intelligence first (PAYE calculations, drought phase monitoring, crop timing), wrap it in AI interfaces second. This is the opposite of the "fine-tune GPT on everything" approach that produces hallucinating agents.
Trust architecture: Every tool in this stack labels DEMO data explicitly, ships with SECURITY.md, uses OIDC publishing. Governance-first development — because the cost of a health AI giving wrong medical advice in Marsabit is not the same as the cost of a chatbot giving wrong movie recommendations in San Francisco.
Swahili-native design: Locale as infrastructure, not localization afterthought.
The Bottleneck Insight
The startup factory analysis notes something important: what survives isn't alumni status. It's whether the person left with a specific bottleneck insight the market now needs.
Tom Mueller left SpaceX having mastered propulsion — the bottleneck. He built Impulse Space targeting orbital transfer, which the emerging satellite economy needed.
My bottleneck insight: M-PESA is Kenya's largest financial institution. It had zero AI interface. KRA processes tens of millions of tax interactions. It had zero AI interface. NDMA monitors drought affecting 4M+ people. It had zero AI interface.
mpesa-mcp, kra-mcp, wapimaji-mcp. That's the bottleneck insight made concrete.
Platform, Not Product
Stripe didn't build the best e-commerce store. They built payment infrastructure. Twilio didn't build the best SMS app. They built communications infrastructure. The value compounds when others build on top.
The 23 MCPs are payment rails for coordination. The apps built on top — the east African health AI, the precision agriculture assistant, the tax compliance chatbot — will be built by people who learned to think about coordination as an infrastructure problem.
That's the East Africa coordination mafia pattern. Not reproducing the same MCP servers. Reproducing the epistemology.
Stack: 23 MCP servers · All MIT · All on PyPI · All Glama-listed
Stack: github.com/gabrielmahia · gabrielmahia.github.io/nairobi-stack
Top comments (0)