Install caldrix

One command installs caldrix and the linker it uses, ld.lld, into ~/.local/bin, verifying both checksums before anything is written and never using sudo. The installer runs on Linux x86_64 with glibc 2.38 or newer and on macOS 14.0 or newer on Apple silicon; the hardware each lane is supported on is listed in index.md.

Install

curl -fsSL https://caldrix.ai/install.sh | sh

The script asks GitHub for the latest release of caldrix-ai/caldrix, downloads the archive for this platform, verifies it, installs the two binaries, then runs caldrix doctor so you see which lanes this machine can use (doctor.md). Releases and their notes are at github.com/caldrix-ai/caldrix/releases; the macOS archive is published from v0.1.1.

The whole script is inside one function called on its last line, so a download cut short runs nothing.

Choose the directory

--prefix DIR installs into DIR instead of ~/.local/bin:

curl -fsSL https://caldrix.ai/install.sh | sh -s -- --prefix "$HOME/bin"

The script creates the directory if needed and never uses sudo; pick a directory you can write to. It takes no other option. When the directory is not on your PATH, the script prints add it to your PATH: and the export PATH="DIR:$PATH" line to paste, DIR being the install directory.

Platforms

Any other platform is refused by name. On Linux the script reads the glibc version from ldd --version and refuses a host below 2.38, or a host whose ldd reports no glibc (musl, for example). The refusal names the container route instead: the CPU image ghcr.io/caldrix-ai/caldrix:latest-cpu (below). On macOS the script reads sw_vers -productVersion and refuses a Mac below macOS 14.0 before downloading anything.

The CPU container image

The CPU image is published as ghcr.io/caldrix-ai/caldrix:0.2.0-cpu (also latest-cpu): linux/amd64, a pinned base image, built from the published release archive. Its entrypoint is caldrix, and its model cache is HF_HOME=/data/huggingface, so a volume there avoids downloading the model again:

docker run -v caldrix-hf:/data/huggingface ghcr.io/caldrix-ai/caldrix:0.2.0-cpu \
  run smollm2-instruct -p "The capital of France is" --max-tokens 8

docker run -v caldrix-hf:/data/huggingface -p 8425:8425 ghcr.io/caldrix-ai/caldrix:0.2.0-cpu \
  serve smollm2-instruct --host 0.0.0.0

--host 0.0.0.0 exposes an unauthenticated, plaintext server on every interface the container is reachable on. Put a gateway with TLS in front of it, or bind to 127.0.0.1 or a private network.

The image is CPU only: its hip lane reads no, and the AMD (ROCm) image is not published yet. Recorded on an MI300X box's host VM: the image pulled by digest printed caldrix 0.2.0, its doctor read the CPU lane usable, and its CPU completion was byte-equal to the installed archive's on the same box.

On Kubernetes, the Helm chart runs this image; see Kubernetes.

What is checked

Each release ships caldrix-<version>-<platform>.tar.gz and its .sha256 file. The archive holds three files: caldrix, ld.lld and SHA256SUMS, the checksums of the two binaries.

1. The archive's checksum is checked against its .sha256 file.
2. After unpacking, both binaries are checked against SHA256SUMS.

If either check fails, the script stops before writing anything to the install directory. Checksums are computed with sha256sum, or shasum -a 256 where that is missing (macOS); with neither, the script refuses.

Each binary is copied beside its destination and renamed over it, so reinstalling while caldrix is running works.

Install with mise

In a project managed by mise, add caldrix to mise.toml:

[tools]
"github:caldrix-ai/caldrix" = "latest"

or pin a release, such as "0.1.1". mise install downloads this platform's release archive from GitHub and puts caldrix and ld.lld on the project's PATH; mise lock records each archive's checksum in mise.lock. mise does not run the installer's platform checks: on Linux below glibc 2.38, or a Mac below macOS 14.0, caldrix fails when it starts instead of being refused before the download.

If the project already has a different ld.lld on its PATH that ours would shadow, expose only caldrix. It still finds its own ld.lld in mise's install directory:

[tools]
"github:caldrix-ai/caldrix" = { version = "latest", filter_bins = "caldrix" }

Uninstall

Remove the two binaries from the install directory:

rm ~/.local/bin/caldrix ~/.local/bin/ld.lld

Downloaded models stay in the Hugging Face cache (models.md).