Your data stays private. Full stop.
Mili bridges the gap between the security of your local workstation and the power of cloud GPUs. The software lives with you; only the prompt travels — encrypted, to sovereign, private GPU infrastructure in the EU — never a US AI service — and never into a log.
Three hops. Nothing stored.
What happens between the moment you press Enter and the moment the answer appears.
Your hardware
Mili runs on your workstation, server or Raspberry Pi. Files, credentials and client data are read here and stay here. Tools run here.
Encrypted transit
Only the prompt and the model's reply cross the wire, over TLS, to an inference endpoint operated by Mili — not a US-based AI service, not a public multi-tenant AI API.
Our own servers, ISO 27001 datacenter
sovereign, private GPUs in the EU, in a certified datacenter outside US jurisdiction, compute the answer, encrypted at rest and in transit. The content of prompts and outputs is never logged, stored, analyzed or used for training.
Functional utilitarian design
We reject superficial features in favor of practical purpose. A single agent reads your files, browses and fills web forms, runs shell commands, queries APIs, and chains those steps into complete jobs — on your machine, under your rules.
Private AI, by construction
Inference runs on sovereign, private GPU infrastructure in the EU in a certified ISO 27001 datacenter — not a US-based AI service, no US provider in the loop — encrypted in transit and at rest. Your data is never used to train any model. Mili keeps no log of your prompts or the model's responses — only metadata (model used, token counts, timestamp) for reliability, security and billing.
What we keep, and what we never see
| Data | Where it lives | Logged by us? |
|---|---|---|
| Your files and documents | Your hardware only | Never |
| Credentials, API keys, tokens | Encrypted vault on your hardware | Never |
| Prompt and model response content | Transits encrypted to our own servers; processed in memory | Never |
| Model name, token counts, timestamp, IP address | Our inference gateway | Yes — reliability, security and billing |
| Account and billing details | Mili (Canada) and Stripe (US) — billing only, never your prompts | Yes — see the privacy policy |
Core infrastructure principles
Proven engineering and practical purpose, for organizations that must automate real work without handing their data to a multi-tenant AI cloud.
- Local first
- The agent, its memory, its tools and its indexes are on your machine. Uninstall it and nothing of yours remains anywhere else.
- Tool-by-tool access
- Every capability — file system, shell, browser, each connected application — is granted individually. The agent cannot touch what you have not enabled.
- Traceable actions
- Each step the agent takes is recorded locally with its reasoning, so you can audit what was done and why.
- Open protocol
- Integrations use the Model Context Protocol. Mili is both an MCP client and an MCP server, so it plugs into the ecosystem instead of locking you in.
- MCP secured with OAuth 2.1
- Both directions authenticate with OAuth 2.1, not just static tokens. As a client, Mili connects to protected MCP servers using authorization-code with PKCE, metadata discovery (RFC 9728 / RFC 8414), dynamic client registration (RFC 7591) and resource indicators (RFC 8707). As a server, Mili is an OAuth 2.1 resource server implementing the MCP authorization spec: external clients — Claude Desktop, Claude.ai custom connectors, ChatGPT — register themselves and sign in through your own identity provider (Microsoft Entra ID, Google Workspace or self-hosted OIDC), with no pre-shared secrets to distribute.
- Privacy by default
- Your data stays under your control; inference runs on our own servers, never on a US AI service; no prompt content is retained; a privacy officer is designated. Locations and processors are listed in the privacy policy.
See it run on your own machine.
Install in one command, then grant access to your first tool.