Telemetry and network
The CLI sends no telemetry. This page lists every network connection it makes.
The amicu CLI sends no telemetry, analytics, crash reports, or update checks. There is no account and no amicu-operated server involved in init, fmt, validate, plan, apply, doctor, or skill export. This was verified against the CLI's code and dependency tree: the only HTTP clients are the ones listed below.
Connections the CLI makes
| When | Destination | What is sent |
|---|---|---|
validate, fmt, init, doctor, skill export | None | Nothing. These commands are offline. Provider region and GPU catalogs are compiled in. |
plan | None, unless state records a remote deployment | Offline for local-only projects. If state holds a remote provider deployment, plan first asks that provider whether it still exists (today only Modal does this; see the provider rows below). |
apply on this machine | 127.0.0.1:<port> | Health probes and the endpoint auth check (GET /v1/models) to the service amicu just started. If you set an api_key, the check sends it as a Bearer header to that loopback address. |
Provider rows: apply, plan and provider commands when a remote provider is involved. Today, Modal: | api.modal.com:443 (gRPC over TLS) | Your Modal token pair for authentication, and the app, image, and function definitions needed to deploy: image build steps, the launch command, GPU and region request, and a Modal Secret holding the endpoint API key (and HF_TOKEN if set). plan and apply also look up whether recorded apps still exist whenever state contains a Modal deployment. |
apply on a remote provider (Modal today) | The deployment's Modal URL | The endpoint auth check, and the proxy token headers (Modal-Key, Modal-Secret). |
amicu modal metrics | The deployment's Modal URL | A GET /metrics with the proxy token headers. Hardware metrics use a read-only probe run inside the container through Modal's API. |
Credentials are sent only to the endpoint they protect. The exposure check does not use environment HTTP proxies and does not follow redirects, so a key is never forwarded elsewhere.
Connections that are not amicu's
Amicu starts runtimes that make their own connections:
- Modal containers download model weights from Hugging Face when they start, and pull the base container image (by default
vllm/vllm-openaiorghcr.io/ggml-org/llama.cpp). If you setHF_TOKEN, it is available to that container. - llama-server and vLLM on your machine are third-party programs. Whatever they fetch (for example a model when given a Hugging Face reference by
extra_args) is their behavior, not amicu's. Amicu itself does not download model files for local deployments. - Clients you point at the endpoint talk to it directly. Amicu is not in the request path and never sees prompts or completions.
Local files
Amicu writes .amicu/ in your project, the service definitions described in This machine, and log files under ~/Library/Logs/amicu/ on macOS. Nothing is uploaded.
The server crates
The repository also contains a separate server (amicu-server) backed by Postgres, used by the hosted product and for self-hosting. It is not part of the CLI workflow documented here.