Custom Runtimes

Add any language by dropping a runtime.yaml and a kernel script in ~/.config/rat/runtimes/<lang>/. No Go code, no recompilation.

runtime.yaml

Here’s the built-in R runtime as an example:

name: r
display: R

detect:
  commands: [Rscript, R]
  env: RAT_R

kernel:
  type: json
  script: kernel.R
  args: ["--no-save", "--no-restore"]

frontend:
  type: repl
  prompt: "r> "

install:
  runtime:
    manager: r
    deps: [jsonlite]
  smoke:
    run: "1 + 1"
    expect: "2"

Two kernel types are supported:

• json: a subprocess speaking the JSON kernel protocol over stdin/stdout
• tmux: an interactive REPL running in a tmux session with a bridge script

Kernel script

For json kernels, the script reads JSON lines from stdin and writes JSON lines to stdout. The wire format is documented in the Kernel Protocol.

Built-in runtimes

The built-in runtimes (R, pi) use the same mechanism — there’s nothing special about them.

Architecture

rat (Go binary)
├── CLI           install, run, look, status, start, stop, ...
├── REPL          rat py → IPython connected to shared kernel via MCP
├── MCP server    rat serve → HTTP or stdio
├── Resolver      project-scoped naming (py@myproject)
└── Kernels
     ├── python   built-in, IPython + jedi
     ├── bash     built-in, direct PTY
     ├── r        generic JSON kernel (runtime.yaml + kernel.R)
     ├── pi       generic tmux kernel (runtime.yaml + bridge.ts)
     └── custom   ~/.config/rat/runtimes/<lang>/

The MCP server is Go. The kernel subprocess runs in the user’s interpreter with access to their packages. Communication is JSON lines over pipes.