io.github.gabrielmahia/offline-mcp

offline-mcp

---

Compatible with claude-sonnet-5 (released 2026-06-30) — Anthropic's most agentic Sonnet yet. Runs multi-step tool chains end-to-end without stopping short. Install: pip install offline-mcp · Use with any MCP client.

---

Local AI inference infrastructure — Ollama wrapper, open weights directory, degraded-mode guide for East Africa.

Why: Never assume OpenAI survives, Anthropic stays accessible, or export controls disappear. This matters more in Africa than anywhere else.

1st world equivalent: Ollama, LLaMA, Mistral local deployment

Why This Exists: Data Sovereignty

"If you take the deal, you're going to be exploited. If you don't take it, you're going to die."
— Frank Ssekamwa, Ugandan digital rights expert

Across the Global South, AI and health data from communities is being extracted, processed abroad, and used to build models whose value flows away from the communities that generated it.

offline-mcp is the sovereignty floor of the East Africa coordination stack.

When this runs on a Raspberry Pi 4 with a 50W solar panel and a 256GB SD card:

• Health data stays in the clinic. Health guidance comes from local models.
• Land records stay in the land office. Queries don't touch foreign servers.
• Civic data stays in the county. AI assistance runs without internet.

No API key. No cloud dependency. No data leaving the community.

The models available via offline-mcp (Llama 3.2, Qwen 2.5) run entirely on device. Community data used to generate AI outputs creates no dataset sent back to model providers.

This is not a privacy feature. It is the architectural foundation of digital independence.

---

Research Foundation

This server implements patterns validated by peer-reviewed research on offline-first AI for bandwidth-constrained environments:

arXiv:2603.03339 (2026) — Offline-First LLM Architecture for Adaptive Learning in Low-Connectivity Environments — confirms that meaningful AI support is achievable with hardware-aware model selection when designed for infrastructure-limited deployment. Key finding: offline-first is a complementary paradigm, not a compromise.

Design principles applied:

• Local-first: all core operations execute without internet connectivity
• Graceful degradation: reduced functionality beats no functionality
• JSONL queue: events accumulated offline sync when connectivity returns
• Hardware-aware: tool selection adapts to available compute

East Africa context: Kenya, Tanzania, Uganda — rural areas with intermittent connectivity represent the primary deployment target. This server is not designed for ideal conditions. It is designed for real ones.

Install

pip install offline-mcp

Tools (6)

Tool	Description
check_ollama_status	Check if Ollama is running locally and list available models
run_local_inference	Run a prompt through a local Ollama model
list_recommended_models	Best open-weight models for East Africa use cases
degraded_mode_guide	4-level degraded mode architecture for offline operation
open_weights_directory	Directory of open-weight models with Africa language support
local_deployment_guide	Deployment guide for laptop, server, Raspberry Pi, Android

Context

Runs on a 50W solar panel + Raspberry Pi 4. Viable for rural Kenya clinics, schools, and community offices.

→ The Nairobi Stack

License

MIT © Gabriel Mahia | contact@aikungfu.dev

Part of the East Africa Coordination Stack

This MCP server is one of 32 tools in 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 32 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 — 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.