Self-hosted AI agents,
on your own hardware.

An AI agent is an LLM-powered process with its own tools, memory, and identity. Self-hosting means running that agent on hardware you own — no cloud control plane, no third-party host, no lock-in. tentaflake is an open-source NixOS template that makes it declarative.

Why self-host your agents?

• Privacy. Your prompts, files, and secrets stay on your machine. You bring your own model keys (or run local models); nothing else phones home.
• Isolation. Each agent is contained — a compromised or misbehaving agent can't reach the others' data.
• Ownership. No per-seat pricing, no vendor shutting the service down, no data hostage. It's MIT licensed.
• Reproducibility. The whole host is declared as code, so rebuilding it — or rolling back — is one command.

Real OS-level isolation

tentaflake gives every agent its own Linux user, its own 0700 state directory, its own container, and its own agenix-encrypted secrets. That's separation at the operating-system level — not just a different folder in the same compose file. When you'd rather share resources, you can run a team of agents in one container via agent profiles: separate personas and sessions, shared skills.

Two runtimes, one host

Hermes is the default runtime — a local assistant tentacle for coding and everyday tasks. ZeroClaw is an autonomous runtime that publishes its own gateway, suited to research and worker-style agents. Mix as many of each as you like on a single machine — one NixOS brain, many tentacles.

Three ways to start

1. Try it live. Boot the live ISO from USB; agents run in RAM and nothing touches disk. Pull the drive and every trace is gone.
2. Install permanently. Boot from USB and let the guided installer deploy NixOS + your agents to disk.
3. Already on NixOS? Add tentaflake as a flake input to your existing config.