Languages

Catalog

The catalog comes from languages.yaml, embedded in the binary (internal/adapters/catalog/languages.yaml). It lists 15 languages (plus the llama stack, below):

Language	API name	Source OCI image	vCPU	RAM	Warmup
Python	python	python:3.12-alpine	1	512 MiB	stdlib compileall
Node.js	nodejs	node:22-alpine	1	512 MiB	—
Go	go	golang:1.25-alpine	2	1024 MiB	GOCACHE populated
Rust	rust	rust:1-alpine	2	1024 MiB	—
Java	java	eclipse-temurin:21-jdk-alpine	2	1024 MiB	—
C	c	alpine:3.22	1	512 MiB	apk add build-base
C++	cpp	alpine:3.22	1	512 MiB	apk add build-base
Ruby	ruby	ruby:3.3-alpine	1	512 MiB	—
PHP	php	php:8.3-cli-alpine	1	512 MiB	—
Perl	perl	alpine:3.22	1	512 MiB	apk add perl
Bash	bash	bash:5.2	1	256 MiB	—
sh	sh	alpine:3.22	1	256 MiB	—
AWK	awk	alpine:3.22	1	256 MiB	—
jq	jq	alpine:3.22	1	256 MiB	apk add jq
SQLite	sqlite	alpine:3.22	1	256 MiB	apk add sqlite

rc languages (or GET /api/v1/languages) shows the same list with ready, which tells whether the image has already been built on this host. A language without a built image is rejected with E006.

Resources

vCPU and RAM are an abuse ceiling, not a reservation. libkrun memory is elastic: the host only pays for the pages the guest touches, and an idle VM measured a 44 MB footprint regardless of the ceiling. Idle vCPUs sleep in the hypervisor.

How code runs

Family	Languages	Execution
interpreted	Python, Node.js, Ruby, PHP, Perl, Bash, sh, AWK, jq	<interpreter> <main file>
compiled	C, C++, Rust	compiles to a binary and runs the binary
JVM	Java	javac and java Main
Go	Go	go run
SQL	SQLite	the code goes through the stdin of sqlite3 :memory:

In every case: the first entry in files is the main file, the working directory lives on the /tmp tmpfs, the VM is fresh and destroyed at the end, and there is no network.

Adding a language

There are two conditions, and both must hold:

1. The language exists in the guest executor (internal/guest/executor.go), with an extension, a command and, if it is compiled, the build command. The executor has entries for more languages than the catalog publishes; if the one you want is not there, it has to be added to the executor (and the guest rebuilt).
2. A new entry in languages.yaml:

  - name: lua
    image: <imagem OCI oficial>
    cpus: 1
    memory_mib: 256
    warmup: ""

Field	Rule
name	exactly the name of the entry in the guest executor
image	a public OCI image for the host architecture; the build pulls it without Docker
cpus, memory_mib	abuse ceiling
warmup	one line of sh, run in a chroot inside the build VM, with network; the result is baked into the image. Nothing from the warmup runs in the job VM

Then:

task build
task image:build -- --stack <nome>
rc run --stack <nome> --code '<programa mínimo>'

The executor's binary must be on the image's PATH or at the absolute path the executor uses. An Alpine-based (musl) image is the tested case; the guest is static and runs on any of them.

The llama stack (GPU)

Besides the 15 languages, the catalog ships the llama stack (fedora:43, 4 vCPU, 4096 MiB, gpu: true): llama.cpp with the Vulkan backend and a Qwen2.5 0.5B model baked into the image. It exists for --gpu; see GPU.

Outside the catalog

• Machines with any image: rc machine create --image <ref> (or image = "..." in the runner.code) builds and boots an OCI image that is not in the catalog. It serves machines, not API jobs.
• Database parsers: the guest executor still has entries for SQL and NoSQL syntax validators, but they have no image in the catalog. See Database parsers.