Hanseatic Alliance for Networked Sovereign Architecture
Make tokens cheap again. The easiest way to run an LLM on your own hardware — a private, OpenAI-compatible endpoint, fully offline. No API keys, no usage bills, no account needed. When you choose to, create a free account, share spare capacity with the community swarm and route your own requests across the network, geo-scoped to your rules.
The same software serves radically different needs — because self-sovereignty, research freedom, and regulatory compliance are not in conflict.
Install H.A.N.S.A. on a gaming PC or an Apple Silicon Mac (tested on one Mac so far). You get a private, OpenAI-compatible endpoint — no API keys, no usage bills, no data leaving your device. Opt into the community swarm when you want more capacity.
The long-term thesis: combining smaller models can outperform single large ones. Mixture of Agents, self-consistency, speculative ensembling — this research exists but lacks a real-world substrate. H.A.N.S.A. aims to provide it.
Data residency is designed to be enforced by the routing layer, not left to a policy checkbox. The coordinator filters peer discovery so a node's candidates are already restricted to its configured boundary — and is designed so the coordinator itself never sees inference data.
The node gives your agent a private model endpoint and a set of tools, so small models do the reading and counting and your main model only sees the answers.
The node finds Claude desktop, Claude Code, opencode and Hermes and adds itself to their MCP config, backing the file up first. Tested live with Claude desktop; the others are in testing. Any other app gets a copy-ready endpoint.
Your agent can search, count, summarize and delegate through the node. Long files and logs are read by small models, on your machine or on peers in your geo scope, and only compact answers come back.
On our own small benchmark (30 Oct 2026, four questions over material too long to read directly, each asked three times), a 35B agent with the node's tools answered every question correctly, against half without them, using about 32 times fewer tokens of its own.
Built on proven open-source primitives. The novel parts are the policy engine, the batch splitter, and the routing layer.
The installer detects your GPU (NVIDIA CUDA, Apple Metal, or CPU only) and suggests the right model tier. One command. Works offline. No account needed to run it.
LiteLLM Proxy exposes a standard OpenAI endpoint on your machine. Any app that calls OpenAI works with H.A.N.S.A. — no code changes.
Enable sharing. Set your geo scope. The coordinator matches you with compatible peers — applying your policy before any peer sees your node.
The batch splitter detects parallel agentic calls and fans them across peers when the local node is loaded. The same primitive enables multi-model ensemble research.
H.A.N.S.A. builds on llama.cpp for local inference (several requests served side by side; Ollama on older systems), LiteLLM Proxy for OpenAI compatibility, and Iroh for encrypted peer transport with NAT traversal. The coordinator handles discovery and geo-policy — and is designed to stay completely off the inference data path.
The novel layer is what we build on top: geo-policy enforcement, latency-aware peer ranking, the agentic batch splitter, signed update distribution, and a public network dashboard.
We publish a full threat model and we don't overclaim. Here's the short version of what each side can and cannot see.
| Claim | Status |
|---|---|
| Running locally is exactly as private as your own machine — it is your own machine. | ✓ True today |
| Your traffic never leaves the geographic boundary you choose. | ✓ Enforced by geo-scoped matching |
| The coordinator is designed never to see your inference content. | ✓ Off the data path by design |
| Every update is signed; nodes reject tampered software. | ✓ Implemented |
| Every answer from a peer carries a receipt signed by the node that served it. | ✓ Implemented · first live run Oct 2026 |
| The serving node never learns who sent a request — prompts travel via relay, anonymized. | 🛠 Rolling out through the alpha |
| Names, passwords, and IDs are stripped on your own machine before a prompt leaves it (PII firewall); prompts flagged sensitive stay home by default, visibly. | 🛠 Rolling out through the alpha · already in the node, being tested between machines |
| A malicious operator could still read what their node processes. | ⚠ True — but once the privacy layer lands, that means an anonymous prompt stripped of names, credentials, and IDs, unlinkable to you. Until then, send only non-sensitive workloads. Node validation & confidential execution come next on the funded roadmap. |
Rule of thumb: until the privacy layer ships, send the public swarm only what you would send a stranger. Confidential material belongs on your own node or a private swarm. Read the full threat model at the institute.
MVP done. Hardening in Q4 2026. Community alpha in Q1 2027.
One command to a private, OpenAI-compatible LLM on your own hardware. The questions take about a minute; the model download depends on your connection.
curl -fsSL https://hansa.institute/get/install.sh | bash
irm https://hansa.institute/get/install.ps1 | iex
Windows / Linux · macOS (Apple Silicon, tested on one Mac) · 8 GB+ RAM · NVIDIA GPU or Apple Silicon recommended · Python 3.11 (the Windows and macOS installers set it up for you; on Linux the installer offers to)
Fully private by default. Joining the network is a separate, optional step and needs an invite.
Running a node on your own machine needs no invite. Joining the public network does: leave your email and we will send an invite when the next group of EU operators is admitted.
No spam. Network invites and major milestones only.
By joining you agree to our privacy policy.