Introduction
What Amicu is, what it is not, and which providers it can deploy to today.
Amicu is a command-line tool for deploying open-weight LLM inference. You describe models, hardware, and deployments in HCL (amicu.hcl). Amicu checks the configuration, shows a plan against recorded state, and applies it by running llama.cpp or vLLM for you.
amicu init # scaffold amicu.hcl and .amicu/
amicu validate # structural, reference, capacity, and policy checks
amicu plan # what apply would do, diffed against state
amicu apply # start, update, or stop deployments after approvalThe workflow is the one Terraform users know: configuration is declarative, every change is a plan you can read first, and state records what was applied.
What it does
- Runs a llama.cpp or vLLM server where you choose: on your machine as a launchd (macOS) or systemd (Linux) user service, or on a remote GPU provider with a pinned region (Modal is supported today; others are in preview). See Providers.
- Checks that the model fits the hardware before anything starts.
- Treats data residency and provider allowlists as resources that
validateandplanenforce. - Keeps credentials out of configuration, plans, state, and command lines. See Secrets.
- Verifies after every deploy that the endpoint rejects anonymous requests (loopback endpoints without an
api_keyskip this check). See Endpoint auth.
What it is not
- Not a hosted service. The CLI runs on your machine and talks only to the providers you deploy to. See Telemetry and network.
- Not a model registry or a gateway. It starts the runtime's own OpenAI-compatible server; clients call that server directly.
- Not a Kubernetes operator. There is no Helm or cluster delivery in the CLI today.
- Not stable. Expect breaking changes to the HCL schema and the CLI.
Status
| Provider | Status |
|---|---|
This machine (local, local_apple, local_cpu, on_prem) | Supported: amicu apply works |
| Remote GPU: Modal | Supported: amicu apply works |
| Hetzner, OVHcloud, Scaleway, RunPod, SSH hosts | Preview: adapter built, not tested live, not wired into apply |
| AWS, Azure, GCP, Nebius | Planned: region and pricing data only |
A deployment on a provider that apply does not support fails with an error. It never falls back to running locally. See Providers.
Where to go next
- Quickstart: install, deploy a model, send a request.
- Concepts: plan, apply and state; resources; residency.
- Configuration: the HCL language and every resource kind.
- Security: secrets, endpoint auth, state, network.
- Reference: CLI commands and error codes.