Mpesa Mcp

mpesa-mcp

MCP server for East African fintech APIs — M-Pesa (Safaricom Daraja) and Africa's Talking

Give your AI agent the ability to trigger M-Pesa payments, check transaction status, send SMS, and top up airtime across 20+ African telecom networks.

[Tests] [PyPI] [License: MIT] [Glama Score]

[Glama] [NSA CSI guidance: self-assessed]

[mpesa-mcp MCP server]

Models

This server does not recommend, certify or benchmark any model, and get_model_hint() is kept only for compatibility (it returns neutral guidance). Use a model that follows tool schemas reliably, rehearse in the sandbox first, and keep a person approving every money-moving or messaging call: in live mode those tools need confirm_send=true and are capped (MPESA_MAX_AMOUNT_KES, default KES 100,000; MPESA_MAX_SMS_RECIPIENTS, default 50).

Why this exists

M-Pesa processes more transactions per day than PayPal does in Africa. Africa's Talking reaches users in 20+ countries on basic phones via SMS and USSD. Neither has an MCP server.

This means every AI agent built today — Claude, GPT, Gemini, or any MCP-compatible runtime — cannot trigger an M-Pesa payment or send a Kiswahili SMS without custom integration work.

mpesa-mcp closes that gap in one pip install.

Tools

Coverage

• M-Pesa: Kenya (Safaricom Daraja v3) — STK Push, C2B, transaction status
• SMS/Airtime: Kenya, Nigeria, Ghana, Tanzania, Uganda, Rwanda, South Africa, and 15+ more via Africa's Talking

Glama (hosted MCP)

mpesa-mcp is available as a hosted MCP server on Glama:

[mpesa-mcp MCP server] [mpesa-mcp score]

Security: how this maps to the NSA's MCP guidance (self-assessed, not certified)

mpesa-mcp was reviewed against NSA CSI U/OO/6030316-26 (May 2026), the NSA Artificial Intelligence Security Center's Cybersecurity Information Sheet on Model Context Protocol security. That sheet is guidance, not a certifiable standard, so the table below is this project's own self-assessment, not an audit or a certification.

See SECURITY.md for how to report a vulnerability.

Reference: NSA CSI_MCP_SECURITY.pdf (May 2026, UNCLASSIFIED)

Install

pip install mpesa-mcp

Or run directly with uvx:

uvx mpesa-mcp

Configuration

Set these environment variables before starting the server:

# M-Pesa (Safaricom Daraja)MPESA_CONSUMER_KEY=your_consumer_keyMPESA_CONSUMER_SECRET=your_consumer_secretMPESA_SHORTCODE=174379               # sandbox test shortcodeMPESA_PASSKEY=your_passkeyMPESA_CALLBACK_URL=https://yourdomain.com/mpesa/callbackMPESA_SANDBOX=true                   # set false for production
# Africa's TalkingAT_USERNAME=sandbox                  # your AT username (sandbox for testing)AT_API_KEY=your_at_api_key

Sandbox credentials

New Daraja account, or setting this up on a new machine? Follow the sandbox setup guide: it records the verified procedure (create the app with only Lipa Na M-Pesa Sandbox ticked, use the published test shortcode and passkey, point MPESA_CALLBACK_URL at a public HTTPS address you can watch) and what a healthy result looks like. Consumer keys cannot be shared, so each developer creates their own free sandbox app.

M-Pesa sandbox: https://developer.safaricom.co.ke — create a free app to get test credentials.

• Test shortcode: 174379
• Test passkey: bfb279f9aa9bdbcf158e97dd71a467cd2e0c893059b10f78e6b72ada1ed2c919

Africa's Talking sandbox: https://account.africastalking.com — use username=sandbox, any API key.

Usage with Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{  "mcpServers": {    "mpesa": {      "command": "uvx",      "args": ["mpesa-mcp"],      "env": {        "MPESA_CONSUMER_KEY": "your_key",        "MPESA_CONSUMER_SECRET": "your_secret",        "MPESA_SHORTCODE": "174379",        "MPESA_PASSKEY": "your_passkey",        "MPESA_CALLBACK_URL": "https://yourdomain.com/mpesa/callback",        "MPESA_SANDBOX": "true",        "AT_USERNAME": "sandbox",        "AT_API_KEY": "your_at_key"      }    }  }}

Usage with Claude Code

claude mcp add mpesa -- uvx mpesa-mcp

Set env vars in your shell before running claude.

Example prompts

Once connected, you can ask your AI agent:

"Send KES 500 STK Push to +254712345678 for order #1234"

"Check if the payment QKL8ABC123 has been received"

"Send an SMS to these 50 farmers with today's maize price: [list]"

"Top up KES 50 airtime for our field agents: [list of numbers]"

Real-world scenarios

Field agent payment dispatch

"Send KES 300 STK Push to each of these 12 field agents for today's data collection: [list]"

The agent triggers 12 sequential STK pushes, tracks each checkout_request_id, and polls for confirmation — without any code from you.

Farmer alert + airtime

"SMS these 200 Garissa farmers that the river is rising. Then top up KES 20 airtime each so they can call in reports."

One prompt → 200 SMS messages and 200 airtime top-ups across Safaricom, Airtel, and Telkom.

Payment reconciliation

"Check whether receipt OKL8M3B2HF was a successful payment and how much it was for"

Useful for support agents using Claude to verify M-Pesa transactions in real time.

Tool annotations

All tools declare MCP tool annotations so clients can gate calls appropriately:

Claude Desktop and other MCP clients will request confirmation before triggering payment, SMS, or airtime operations.

Server discovery

Capabilities are advertised via .well-known/mcp.json — the emerging MCP Server Cards standard. Registries and browsers can index this server's tools without connecting to it.

# Check capabilitiescurl https://raw.githubusercontent.com/gabrielmahia/mpesa-mcp/main/.well-known/mcp.json

Testing and accuracy

These tests check this server's behaviour. This README makes no claim about how accurate any model is when it uses this server.

mpesa-mcp is tested against all three Kenyan phone number formats, boundary amount values, and missing optional fields:

pytest tests/ -v  # run full suitepytest tests/test_phone_formats.py  # format normalizationpytest tests/test_boundary_amounts.py  # min/max amount edge cases

Write operations (STK push, SMS, airtime) have explicit validation before any API call is made.

Ecosystem context — Mojaloop + MCP

Mojaloop (funded by the Gates Foundation) handles payment interoperability — connecting banks, mobile money wallets, and merchants across DFSPs in East Africa and beyond.

mpesa-mcp handles the AI agent tooling layer — enabling AI coding assistants to trigger and query M-Pesa payments programmatically.

These are complementary:

• Mojaloop: the interoperability rails between financial providers
• mpesa-mcp: the MCP interface layer that connects AI agents to those rails

See the Mojaloop documentation contribution for more on this pattern.

MCP vs A2A — two different protocols

mpesa-mcp implements MCP (Model Context Protocol) — how an AI agent talks to tools.

There is a complementary protocol, A2A (Agent-to-Agent), which handles how agents talk to each other. They solve different problems and work together:

• MCP: Your AI agent → mpesa-mcp → Daraja API / Africa's Talking
• A2A: Orchestrator agent ↔ payment sub-agent ↔ notification sub-agent

For most integrations you only need MCP. A2A becomes relevant when you're building multi-agent systems where a payment workflow coordinates with other specialized agents.

Development

git clone https://github.com/gabrielmahia/mpesa-mcpcd mpesa-mcppip install -e ".[dev]"pytest tests/ -v

Security

Do not commit API keys. Use environment variables or a secrets manager.
Report vulnerabilities to: [email protected]

Research Context

Swahili and translation (arXiv:2509.04516, 2025): in a small news-classification experiment, a Swahili-native BERT model made 0.36% errors against 1.47% for an English model applied to machine-translated Swahili, so native beat translate-then-classify on that task. This server's tool descriptions are written so they can be read in Swahili as well as English; no effect of that has been measured here.

MCP security research (arXiv:2603.18063, arXiv:2603.21642, 2026): Prompt injection via tool descriptions is the primary MCP attack vector. mpesa-mcp mitigates this through static, versioned tool descriptions and strict input validation.

Related infrastructure:

• wapimaji-mcp — Kenya drought-response MCP (simulated data)
• civic-agent-kit — Kenya civic data MCP
• swahili-health-mcp — Kenya DHS 2022 county health indicators MCP
• kenya-legal-rag — Kenya legal corpus MCP
• Full portfolio: gabrielmahia.github.io

License

MIT — © 2026 Gabriel Mahia

Stay updated

Get notified of new releases and East African API developments:

Or watch this repo on GitHub for release notifications.

Sibling packages

Related packages

All MIT · All part of the East African civic AI stack

Full portfolio: gabrielmahia.github.io

Part of the East Africa Coordination Stack

This MCP server is part of the Kenya coordination infrastructure. Connect it to africa-coord-bus — the coordination event bus that routes signals between domains automatically.

pip install africa-coord-bus

All servers: pypi.org/user/gmahia Live demo: coord-cascade-demo

IP & Collaboration

MIT licensed. Feedback via GitHub Issues only — pull requests are not accepted. Demo data is labeled DEMO and is not suitable for operational decisions. Full policy: docs/architecture/IP_POLICY.md. Security reports: see SECURITY.md.

Part of the East Africa coordination stack

• Install & run: pip install reli-cli && reli list — the 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.

Safety limits for tools that send money out (0.3.0, breaking in live mode)

A language model can be talked into paying anyone any amount: prompt injection against payment tools is a documented attack on agent payment protocols (AP2, x402). So in live mode (MPESA_SANDBOX=false) the eight outflow tools (mpesa_b2c, mpesa_business_paybill, mpesa_business_buygoods, mpesa_business_pochi, mpesa_reversal, mpesa_tax_remittance, mpesa_b2b_express_checkout, airtime_send) are bounded:

• Confirmation: they refuse unless called with confirm_send=true. The agent must show the user the recipient and amount and set it only after explicit approval.
• Cap: amounts above MPESA_MAX_AMOUNT_KES (default 20,000; 0 = no cap) are refused.
• Session budget: one server process may authorise at most MPESA_SESSION_BUDGET_KES in total (default 50,000; 0 = no budget). The cap bounds one payment; the budget bounds the sum, so an agent cannot repeat a capped payment without limit. It is counted when a payment is authorised and is not refunded if the call later fails (a timed-out call may still have moved money), and it resets only when the server restarts. Only the operator, not the model, can change it. If you legitimately move more than 300,000 KES in one session, set it higher.
• The person confirms, not the model (0.5.0): for the eight money-out tools in live mode the server itself asks you, through your MCP client, to approve the exact payment (recipient in full, amount, reference) before anything runs, using MCP elicitation in both protocol eras. The model cannot answer for you: it can say confirm_send=true, and you can still decline. Your approval is bound to that exact call (a different amount or recipient needs a new approval), works once, and expires after five minutes. Payments the server would refuse anyway (over the cap, over the budget, not on the allowlist) never reach you. MPESA_CONFIRM_MODE: auto (default: ask when your client can answer; otherwise fall back to the confirm_send parameter), elicit (strict: refuse when your client cannot show a prompt) or param (the old behaviour). The sandbox is never prompted. Not every client supports elicitation yet; if yours does not, use elicit mode to refuse instead of falling back, and keep your client's own per-call approval on.
• Allowed recipients: set MPESA_ALLOWED_RECIPIENTS (comma-separated phone numbers, paybills and tills) and the money-out tools will pay only those recipients; anything else is refused. Phone numbers match in any Kenyan format. Unset means no restriction, and for production use a list is recommended: scoping who can be paid is the same idea as a merchant-scoped payment token or a named-merchant reservation. Reversal and tax remittance have no free recipient, so the list does not apply to them.
• Session report: the mpesa_session_report tool (read-only) returns the limits, what has been spent and what is left, and the last ten authorisations or refusals with recipients masked, from the server's own record. Use it to check the budget before a payment, and to verify what the server actually authorised instead of relying on what a model says it did.

Bound the loss outside the tool as well

These guards run inside the tool, the layer an agent controls. Two checks sit outside it, and you should use both, and keep the limits small:

1. confirm_send is supplied by the model itself, so on its own it is not an independent check. Since 0.5.0 the server asks you directly through your client (see above); where your client cannot show that prompt, keep your MCP client's own per-call approval on for the eight money-out tools and never set them to "always allow".
2. Bound the loss in the payment rail. Pay out from a dedicated business shortcode or working account that holds only what you are prepared to lose, and use whatever initiator limits or approval (maker-checker) controls your M-PESA organisation account offers; this server cannot see them, so check Safaricom's current documentation.

This is the layered design that India's agentic-payment pilots (a pre-funded, merchant-scoped reservation and delegated limits enforced by the network) and Alipay's agent payments (a user-set threshold below which an agent may pay and above which a person confirms, a way to review and revoke an agent's authority, and a loss guarantee) arrived at independently.

3. Start small. The defaults here (20,000 KES per payment, 50,000 per session, lowered in 0.5.0 from 100,000 and 300,000) are ceilings, not recommendations. Brazilian banks that let customers pay by voice or text through an AI assistant on WhatsApp start at roughly R$200 to R$800 per payment, ask for a confirmation on each one, and send anything larger to the bank's own app: the same order of magnitude as the defaults here. Set MPESA_MAX_AMOUNT_KES and MPESA_SESSION_BUDGET_KES to the smallest figures you actually need, and raise them only when you have a reason.

• Opt-out: MPESA_REQUIRE_CONFIRMATION=false removes the confirmation requirement for fully automated setups you control. Do that only deliberately.
• The sandbox is never restricted. Customer-approved flows (STK Push, Ratiba, Bill Manager) are unchanged: the customer approves on their handset.