Decision Intelligence

The largest open-source MCP coordination infrastructure for the Global South.
31 servers. Local-first. Data-sovereign. MIT licensed.

The engineering guide for building software products in East Africa — and everywhere institutions are under-resourced.

License: CC BY 4.0 PRs Welcome

A practical, opinionated reference for developers — in Kenya, the diaspora, or anywhere building for East African users. Covers APIs, infrastructure choices, UX patterns, regulatory context, and the cultural nuances that documentation never mentions.

---

What Gets Built On This

These are not hypothetical. They are the trains for which these rails were built.

---

Scenario 1: The rural health worker

A community health worker in Kisumu runs afya-mcp + offline-mcp on a tablet with no internet. She asks: "What is the correct malaria treatment for a 4-year-old weighing 15kg?" The query runs against a local Llama model. No data leaves the tablet. The answer arrives in Swahili in under 3 seconds. Rails used: afya-mcp, offline-mcp, tafsiri-mcp

---

Scenario 2: The smallholder farmer

A maize farmer in Nakuru wants to know whether to sell now or hold for two weeks. She opens WhatsApp, sends a voice note to a bot powered by soko-mcp. The bot returns current prices across eight East African markets, trend direction, and a sell/hold recommendation with the reasoning shown. Rails used: soko-mcp, kilimo-mcp

---

Scenario 3: The Kenyan diaspora member in Virginia

A Kenyan living in Manassas, Virginia discovers that a caretaker has encroached on his father's land in Kiambu. He needs to understand his legal options, initiate a dispute, and manage the process remotely. He uses ardhi-mcp + familia-mcp + fomu-mcp to understand the legal path, draft the initial letter, and track required documents — without hiring a Kenyan lawyer for the preliminary steps. Rails used: ardhi-mcp, familia-mcp, fomu-mcp, diaspora-mcp

---

Scenario 4: The first-time SACCO member

A 23-year-old boda boda rider in Nairobi wants to join a SACCO but doesn't know which one fits his income profile, what the obligations are, or whether the institution is legitimate. jumuia-mcp walks him through SACCO types, finds SASRA-regulated options in his county, and explains what joining actually commits him to. In Swahili. On his phone. Rails used: jumuia-mcp, sifa-mcp, tafsiri-mcp

---

Scenario 5: The county government AI deployment

A county government in Western Kenya wants to deploy an AI assistant for citizen services — permit status, budget enquiries, ward information. They deploy the SII Stack on a local server. Citizen queries hit the sovereign tier first. Only complex reasoning escalates to cloud inference. Citizen data never leaves the county server. Rails used: county-mcp, fomu-mcp, habari-mcp, SII Stack sovereign tier

---

Each scenario above can be built today. The servers are live on PyPI. The SII Stack is open-source. The architecture is documented. What's needed is someone who wants to run the train.

→ SII Stack · Browse all 31 servers · PyPI

Why this exists

Western developer documentation assumes broadband, credit cards, app stores, and English-only users. East African users may have any combination of: feature phones, sporadic 2G, M-Pesa (no bank account), Kiswahili or Sheng as primary language, prepaid data, and screen sizes that haven't been in a Western product review since 2014.

This guide is built from real integration work across three production systems — Catholic Network Tools, OpenResilience, and RemitLens — plus field experience in Nairobi, Mombasa, and across Kenya's 47 counties.

---

Guides

Payments

• M-Pesa Integration — STK Push, B2C, webhooks, sandbox-to-live
• Remittance APIs — Wise, Remitly, Sendwave: what's available, what's not

Communication

• USSD Patterns — menus, sessions, Africa's Talking
• SMS Infrastructure — Africa's Talking, shortcodes, DLT registration

Data & Live APIs

• Live Data APIs — 11 free keyless APIs confirmed working in Kenya: Open-Meteo, NDMA, COB, World Bank, WFP/HDX, LSK/FIDA/Judiciary RSS, open.er-api.com, Yahoo Finance

Design

• Mobile-First UX for Kenya — offline, low-data, feature phone realities
• Kiswahili in Your Product — translation, tone, Sheng, code-switching

Infrastructure

• Deploying in East Africa — latency, CDN, AWS vs local hosting
• Kenya Data Protection Act — ODPC compliance, consent, cross-border data

Finance & Regulatory

• Regulatory Landscape — CBK, CMA, Communications Authority
• Chama Digitisation — domain model, M-Pesa integration, ROSCAs

---

Why This Exists

"Build Rails, Not Trains."

TCP/IP didn't control the internet — it enabled it. Railroads didn't control freight — they enabled an economy. M-Pesa didn't control money — it enabled transactions at scale.

The 31 MCP servers in this stack are rails.

They give AI agents structured, authenticated, locally-processed access to: the M-PESA API, Kenya's 47-county government layer, NDMA drought data, land title systems, health infrastructure, education records, and civic institutions.

Any developer — anywhere — can pip install mpesa-mcp and give an AI agent the ability to trigger a mobile payment in Kenya in under 120 seconds.

That's the rail. What you build on it is the train.

---

Structural Absences, Not Competitive Gaps

These servers weren't built because no one had thought of payments, drought data, or land records. They were built because the engineer with the right intersection of skills hadn't built them yet.

The intersection: Kenyan institutional knowledge + diaspora perspective + AI infrastructure fluency.

That intersection doesn't exist in San Francisco or London. It exists here.

---

The Stewardship Principle

Every decision in this stack was made with one question: "What should I preserve, improve, and hand forward?"

• MIT license: so others can build on it without asking permission
• Local-first architecture: so communities steward their own data
• Human-in-the-loop: so AI assists, never decides unilaterally
• Offline sovereign tier: so the stack survives when the internet doesn't

This is not a product. It is infrastructure. Infrastructure is built for the people who come after you.

---

Built for Africa. Deployable Anywhere.

The architecture is continent-generic. The data is local.

Every server in this stack separates infrastructure (tools, schemas, routing logic) from data (the Kenya-specific records, county budgets, NDMA classifications).

Replace the data layer with Uganda's NSSF records, Nigeria's CAC registry, Ghana's GhIPSS payment rails, or Tanzania's BRELA business registration — and the same architecture works without changing a line of application code.

Current depth: Kenya (31 servers)

Kenya is the starting point because it has the most developed data infrastructure for this kind of tooling: M-Pesa, a mature Daraja API, NDMA drought data, a 47-county government layer with structured data, and an active tech ecosystem.

Expansion logic: depth before breadth

The right expansion sequence is not "build one thin server for every African country." It is "build deep coverage in one country, then lift and replicate."

The next depth targets, in order:

1. Tanzania — 60M people, Swahili national language, fewer legacy constraints, quieter governance
2. Uganda — 48M people, mobile money penetration, East African Community integration
3. Nigeria — 220M people, largest African economy, fintech ecosystem (OPay, Flutterwave, Paystack)
4. Ghana — 33M people, GhIPSS, stable institutions, English-speaking

Each is a separate data layer on the same rail architecture.

What "Africa-wide" actually means

Swahili covers East Africa and large parts of Central Africa (including DRC, Rwanda, Burundi). tafsiri-mcp and swahili-civic-nlp are already cross-border tools. sifa-mcp (portable reputation) and soko-mcp (market prices) have no inherent Kenya boundary — the schema works for any East African market.

---

📊 Portfolio Stats — June 2026

Platform	Status
PyPI packages	31 live (v0.1.1 · v0.1.9 for mpesa-mcp)
Glama directory	31 indexed
Smithery.ai	Indexed (3+ confirmed)
awesome-mcp-servers	mpesa-mcp listed · PR #8363 open
HuggingFace datasets	5 datasets · 246 total downloads
Dev.to articles	15 published · 244+ views
DPGA submissions	2 under review · 1 pending completion

Contributing

This is a living document. If you've built something in East Africa and learned something the hard way, open a PR. See CONTRIBUTING.md.

---

Maintained by Gabriel Mahia. Kenya × USA.

MCP ecosystem updates (2026)

Key developments relevant to East African developers building on mpesa-mcp:

Tool annotations (spec 2025-03-26)

All MCP tools should declare behavioral hints. For payment/SMS servers this is critical:

@mcp.tool(annotations={
    'title': 'M-Pesa STK Push',
    'readOnlyHint': False,      # modifies state (moves money)
    'destructiveHint': True,    # irreversible financial operation
    'idempotentHint': False,    # each call creates a new transaction
    'openWorldHint': True,      # reaches Safaricom's external API
})
def mpesa_stk_push(phone: str, amount: int, ...):

Without these, Claude Desktop and ChatGPT treat all tools as potentially destructive and may add unnecessary friction on read operations.

Server Cards / .well-known (roadmap Q2 2026)

Servers can advertise capabilities via .well-known/mcp.json — a static JSON file at the repo root that registries can index without connecting to your server. Add this to any MCP server you publish.

MCP accuracy benchmark

CData's 2026 benchmark found MCP servers accurate 60–75% on complex queries. Write operations (payments, SMS) are the highest-risk category — test for:

• All phone number formats: 0712345678, 254712345678, +254712345678
• Boundary amounts (minimum 1 KES, maximum per API limit)
• Missing optional fields (reference, description)

Streamlit 1.55.0 (March 2026)

Key changes for East African app builders:

• st.metric now has delta_description parameter — add context like "vs last month"
• st.tabs supports on_change — dynamic loading of heavy data only when tab is opened
• st.image has link parameter — clickable images for maps and charts
• Widget binding (bind parameter) — sync widget state with URL query params for shareable links

AGENTS.md — tell AI coding agents how to work with your repo

AGENTS.md is a spec by OpenAI, now part of the Agentic AI Foundation (Linux Foundation). It's been adopted by 60,000+ repos and is supported by Cursor, GitHub Copilot, Claude Code, Devin, Gemini CLI, and VS Code.

Add it to any repo you want AI coding tools to understand correctly:

# AGENTS.md

## What this is
[Brief description]

## Architecture
[Directory structure]

## Critical rules
[Things agents must not change or break]

## Running locally
[Commands]

All repos in the gabriel Mahia portfolio now have AGENTS.md files.

MCP vs A2A — know the difference

Both are 2025-2026 protocol standards under the Linux Foundation:

	MCP	A2A
What	Agent ↔ Tools	Agent ↔ Agent
Created by	Anthropic	Google
Use case	Connect Claude to M-Pesa	Orchestrate multiple agents
East Africa relevance	mpesa-mcp, WapiMaji SMS	Multi-agent payment workflows

mpesa-mcp is MCP only. For multi-agent architectures (payment + notification + audit), A2A would coordinate the agents while each agent uses MCP to reach its tools.

MCP Dev Summit — April 2-3 2026, New York

The official MCP community event, now under the Agentic AI Foundation. CFP is open. If you're building on MCP for African markets, this is the venue to present.

events.linuxfoundation.org/mcp-dev-summit-north-america

Community presence and contributions

Where to find this work

Platform	Link	What's there
GitHub	gabrielmahia	15 repos, 8+ packages
PyPI	mpesa-mcp	MCP server for M-Pesa + AT
MCP Registry	io.github.gabrielmahia/mpesa-mcp	Official listing
Portfolio	gabrielmahia.github.io	All 13 apps
Engineering blog	aikungfu.dev	Technical writing

Open source contributions

Org	Issue/PR	What
Gates Foundation / Mojaloop	#553	MCP + payment rails integration guide
OCHA / HDX	#194	Kenya food security data recipe
Andrew Ng / context-hub	PR #52	Africa's Talking + M-Pesa docs
mySociety / theyworkforyou	#1998	Kenya parliamentary data tools
Code for Africa / openAFRICA	#55	Kenya civic tools listing
Open Knowledge Foundation	#407	Kenya open data portals
Open Contracting	#1745	Kenya procurement transparency
Ushahidi	#5048	Kenya crisis mapping tools
Open Data Day	#425	Kenya civic tools showcase
Africa's Talking SDK	#72	mpesa-mcp community project

Part of the East Africa coordination stack

• Install & run: pip install reli-cli && reli list — 33 MCP servers on the official MCP Registry under io.github.gabrielmahia
• Evaluate any model on Swahili agent tasks: kipimo · dataset · leaderboard
• Coordinate across servers: africa-coord-bus — offline-first event bus with a built-in Kenya routing table
• Datasets: huggingface.co/gmahia · Docs hub: nairobi-stack

Model-agnostic by design: closed APIs, open-weight models, and small distilled models are all first-class citizens.