Singapore's Minister for Foreign Affairs, Dr Vivian Balakrishnan, used his keynote at AI Engineer Singapore on 16 May 2026 to unveil NanoClaw — a self-hosted personal AI assistant he built around Anthropic's Claude, running entirely on a Raspberry Pi 5 sitting on his desk. The full architecture and code were quietly published on GitHub on 27 April 2026 and have since been reverse-engineered, replicated, and dissected by the global developer community.
It is one of the rare cases of a sitting Cabinet minister shipping a working AI system in public — and the technical decisions behind it have become a reference architecture for privacy-first personal agents.
Key Highlights
- NanoClaw runs on a Raspberry Pi 5 with 8 GB of RAM and connects to WhatsApp, Gmail, Telegram, Slack, and Discord through open-source libraries
- Voice notes are transcribed on-device using whisper.cpp, and semantic search runs locally via Ollama with the nomic-embed-text model, so document content never leaves the network
- A three-tier memory inspired by Andrej Karpathy's "LLM Wiki" pattern turns the minister's speeches and articles into a structured knowledge graph that he reads back through Obsidian, synced via iCloud
- The complete blueprint is public on GitHub Gist, and Balakrishnan told the audience he "dares not switch it off"
What NanoClaw Actually Does
NanoClaw is a lightweight, containerised agent framework built on the Claude Agent SDK. Each chat group runs in its own isolated Docker container, invisible to the others — a design choice that makes the system safe to expose to multiple messaging channels at once.
WhatsApp integration uses Baileys, an open-source implementation of the WhatsApp protocol that avoids commercial APIs. Voice notes arriving on any channel are pulled into a local pipeline, transcribed with whisper.cpp, and routed to Claude with the relevant memory context. Outbound replies can be voice, text, or summaries.
The most distinctive part is the memory layer, which Balakrishnan calls mnemon. It is a graph-based system that stores entities, causal links, temporal relations, and semantic embeddings in SQLite. A nomic-embed-text vector model — only 274 MB — handles semantic search locally, while an LLM-supervised process compiles the raw material into a human-readable wiki organised by entity, concept, and timeline.
Why It Matters
In his speech, Balakrishnan made three arguments that go well beyond his personal setup.
First, personal understanding cannot be outsourced. "The one thing which you cannot outsource is your personal understanding," he said, framing his hands-on coding as a governance posture: "You cannot govern a technology that you have only been briefed on."
Second, the real economic value of AI will come from the ground up. "Real value for the economy and society is created at the ground level — workflow by workflow, sector by sector." Teachers, lawyers, doctors, and managers assembling tools for their own work, he argued, will produce more impact than any top-down national platform.
Third, and most quoted by developers afterwards: "Tools matter more than models." The choice of frontier model is now almost a commodity decision; what differentiates a useful agent from a toy is the surrounding stack — memory, integrations, sandboxing, observability, and the discipline to keep the codebase small enough that one person can maintain it.
Background
Balakrishnan, a trained ophthalmologist before entering politics, has been in Singapore's Cabinet for over two decades and has served as Foreign Minister since 2015. His monthly travel calendar regularly spans a dozen countries, and he has described NanoClaw as a way to compress the preparation, recall, and synthesis work that the role demands — drafting first cuts of speeches, summarising briefings, retrieving facts from his own archive of speeches during live discussions.
The architecture is also a deliberate sovereignty statement. By running embeddings and transcription on-device, the system keeps sensitive diplomatic material off third-party servers. API access to external services is brokered through credential proxies, and filesystem mounts are whitelisted to block sensitive directories such as SSH and cloud keys.
What's Next
NanoClaw's public release has already triggered a small ecosystem of forks and adaptations. Developers across the MENA region, India, and Latin America have begun publishing their own variants, often swapping Claude for other frontier models or adding regional messaging channels. For governments and enterprises evaluating personal AI deployments, NanoClaw offers a ready-made template that is small enough to audit line by line — a feature that almost no commercial agent platform can match today.
For the rest of us, the more durable lesson is the one Balakrishnan delivered to a room of engineers: the barriers to building useful AI have already collapsed. The remaining question is who picks up the tools and starts.
Source: Ministry of Foreign Affairs of Singapore — Speech at AI Engineer Singapore