your data.
qwen2.5:1.5b on Hetzner EU. No signup. 5 sec.
Cloud AI leaks your most sensitive data.
Data leaves your network
Every ChatGPT query, every Claude prompt, every Copilot completion sends your business data to US servers. For regulated industries, this is non-negotiable risk.
GDPR + EU AI Act
Europe's regulatory framework makes sending personal data to cloud LLMs legally perilous. Article 30 requires full audit trails of every data transfer — something SaaS AI tools can't provide.
Vendor lock-in & trust
Third-party AI providers change terms, raise prices, and train on your data. Sovereign AI means you own the stack — inference, memory, audit logs.
Choose your deployment model.
Both products solve the same problem — privacy-first AI — through different architectures. Pick the one that fits your scale, compliance needs, and team.
Magi Edge
- Fractal node mesh — same binary at every layer
- PII scan + redaction on every escalation
- Local-first: Ollama 7B on your laptop
- GDPR consent gate before any data leaves
- Zero cloud dependencies
- MIT License — fully open, fully inspectable
Magi Cloud
- Managed on-prem / VPC deployment
- Team management + admin dashboard
- Compliance-oriented audit trails
- Built on the same privacy-first core as Edge
Built for industries where data matters.
These are the scenarios where cloud AI is off the table — and where Magi Elior thrives.
Client data that can't leave the bank.
Wealth managers analyzing client portfolios. Compliance officers reviewing transactions. Risk teams summarizing filings. All with PII that legally cannot be sent to third-party AI.
Patient records, never in the cloud.
Clinicians summarizing patient charts. Medical research across cohort data. Insurance claims analysis with diagnosis codes. All HIPAA and EU patient data protection territory.
Client privilege, protected by architecture.
Law firms reviewing contracts with client names. M&A teams summarizing deal documents. Litigation support with case strategy.
Sovereign AI for governments.
Ministries processing citizen data. Defense organizations analyzing intelligence. Regulators reviewing filings. All with strict data-residency requirements.
The Fractal Mesh
Layer 0: ROOT a server you run — largest model, final fallback ↑ anonymized escalation (PII already stripped) Layer 1: ENTERPRISE your on-prem server — mid-size model ↑ PII stripped, consent verified Layer 2: DEPARTMENT team server — smaller model (optional) ↑ first escalation target Layer 3: EMPLOYEE your laptop — small model via Ollama Model size at each layer is your deployment choice. The public demo runs qwen2.5:1.5b on a Hetzner EU box. CapabilityVector (6D) decides at each layer: Can I handle this locally? Yes → stay. No → escalate. Escalation gated by: PII scan → consent check → audit log. In Edge mode, raw data never leaves your device.
Engineered for European data sovereignty.
Built in Berlin. Hosted on EU infrastructure (Hetzner, Nürnberg). Open source at the core. Designed from day one around GDPR, the EU AI Act, and the regulatory reality of European business.
Edge vs Cloud
| Dimension | Magi Edge (available today) | Magi Cloud (in development) |
|---|---|---|
| Architecture | Fractal mesh (N nodes) | Managed deployment of the same core |
| Install | pip install + Ollama | Managed — not yet available |
| Inference | Ollama-first (local) | Planned: Ollama / vLLM / LM Studio |
| Privacy model | PII-gate on every hop | Planned: on-prem deployment |
| Auth | JWT + bcrypt + 2FA (TOTP) + mTLS | Planned: SSO / SAML |
| Compliance | GDPR (consent gate, Art. 15/17/20/30 endpoints) | Planned |
| Agents | ReAct engine + tools | Planned |
| Billing | Not enabled on the public demo | Planned |
| Frontend | Minimal chat UI (zero-build) | Planned dashboard |
| Plugin system | Local plugins (example set, bundled) | Planned |
| Best for | Data sovereignty, edge deployment | Teams wanting managed on-prem |
Frequently Asked
Which product should I choose?
Magi Edge is what exists today — self-hosted, privacy-first, available now. Magi Cloud (a managed deployment of the same core) is in development; if you want managed on-prem rather than running it yourself, join the waitlist.
Do I need my own GPU?
No GPU is required. Ollama runs small models (1.5B–8B) on any modern laptop, and the public demo runs qwen2.5:1.5b on a CPU-only Hetzner box. Bigger models mean better answers — that is a hardware choice you make per node, not a product limit.
How does escalation work without leaking data?
Every escalation passes through a PII scan covering 6 entity types (email, phone, SSN, credit card, IBAN, IPv4) that redacts sensitive data before sending. The user must grant explicit GDPR consent for escalation to happen at all. Every escalation is logged with a SHA-256 task hash for audit compliance. Note: redaction is pseudonymisation — it reduces exposure, it is not a legal basis for transferring personal data.
Is this really open source?
Magi Edge is MIT-licensed — fully open, you can fork it, self-host it, inspect every line. Magi Cloud is BUSL-1.1 — the source is available for review and non-commercial use, with commercial licenses available. Both are developed in the open on GitHub.
Can I deploy this in my VPC / on-prem?
Yes, that's the whole point. Magi Edge ships with Docker Compose for air-gapped deployments and also works as a native Python install. No cloud services required at any layer.
Where is Magi Elior hosted?
The public demo runs on Hetzner in Nürnberg (Germany, EU). But that's just one deployment — the whole point of Magi is that you run it where your data already lives. We don't want your data.
Install in under a minute.
No SaaS signup. No credit card. Run locally, in your VPC, or on-prem — your call.
pip install "git+https://github.com/konscious0beast/MagiElior.git"
# not on PyPI yet (release follows) — while the repo
# is private: git clone … && pip install -e .
ollama pull llama3.2
magi-elior start --mode root
# root = standalone; for mesh see docs
Join the waitlist →
# Managed on-prem deployment — not yet available
Your data. Your rules.
Try it right now. No signup. No cloud dependency.