Terminalwire (Elixir)
Ship a CLI for your web app. No API required.
Terminalwire streams a command-line app straight from your Phoenix/Plug server to your users' machines over a single WebSocket. You write your CLI in your app — calling your contexts, Ecto, and business logic directly — and it runs on the user's workstation with their terminal, files, and browser.
A CLI usually costs you three things to build: a REST API to back it, an SDK
or client binary to ship, and a release-and-auto-update pipeline to keep that
client current. Terminalwire is all three. Users install one small, self-updating
client with a single curl … | bash; you ship features by deploying your server.
Terminalwire client ⇄ WebSocket endpoint ⇄ Terminalwire.WebSock
⇄ Server.Session (protocol)
⇄ Server.Context ⇄ your CLI handler
Why this is nice
- No API to build or version. Your CLI calls your app's code directly — no serializers, no SDK, no client/server version skew.
- Nothing to distribute or update. Users install one small client
(
curl <app>.terminalwire.sh | bash) that self-updates through a signed channel. You ship a change by deploying your server — no per-release client build, no app-store round trip. - It feels local. Output streams in real time, prompts and passwords work,
it's color/TTY-aware, resizes with the window,
Ctrl-Cinterrupts the server-side command, and you can pipe into it (cat data.csv | your-app import). - Secure by construction. The client is the trust boundary: the server requests access to a file/env var/the browser and the client enforces a per-app entitlement policy. Your server never touches the user's machine.
- One BEAM process per session. Each connection is a supervised process; the CLI handler runs in its own task. Natural fit for Phoenix.
- Same protocol, any client. This server speaks the exact wire protocol the Go client and the Ruby server do — proven by a shared conformance corpus.
Install
def deps do
[
{:terminalwire, "~> 0.1"},
{:websock_adapter, "~> 0.5"} # to upgrade a Plug/Phoenix conn to a socket
]
end
Use
Define your CLI as a module. Public functions are commands, their parameters are
the command's arguments, and @desc is the help text — like Ruby's Thor:
defmodule MyApp.CLI do
use Terminalwire.CLI, name: "my-app"
@desc "Greet someone by name"
def hello(name) do
puts("Hello, #{name}!")
end
@desc "Deploy to an environment"
def deploy(env) do
if String.trim(gets("Deploy to #{env}? [y/N] ")) == "y" do
puts("Deploying #{env}…") # call your app's code right here
else
puts("Aborted")
end
end
end
Mount it on a WebSocket route — use generated run/1 for you:
# Plug / Bandit / Cowboy
WebSockAdapter.upgrade(conn, Terminalwire.WebSock, [handler: &MyApp.CLI.run/1], [])
That's a working CLI: my-app hello Ada runs hello("Ada"), my-app deploy staging
runs deploy("staging"), and my-app (or my-app help) prints a generated command
list. Inside a command, puts/print/warn/gets/read_secret/env talk to the
user's terminal; context/0 reaches files, the browser, and the rest.
Want flags, options, or your own parsing? Terminalwire.CLI is a thin layer over a
plain run/1 handler — drop down to it and use any parser. That's the next section.
The handler API
Terminalwire.CLI is a thin layer over a plain handler: a one-argument function
that takes a Terminalwire.Server.Context. Use it directly when you want full
control over parsing. It's called once the handshake completes, in its own BEAM task
whose group leader is a Terminalwire IO device, so plain IO.puts/IO.gets,
IO.ANSI, and any library that writes to standard IO (like
Owl) stream straight to the user's terminal. The Context
covers everything that isn't standard IO: args, prompts, the client's terminal,
files, env, the browser.
def run(ctx) do
case Context.args(ctx) do
["deploy", env] -> deploy(ctx, env)
_ -> Context.warn(ctx, "unknown command"); 1
end
end
The Context API
| args | Context.args(ctx) → the argv list you parse |
| stdout | Context.puts/print — or just IO.puts / Owl.IO.puts (group leader) |
| stderr | Context.warn(ctx, msg) (see the stderr rule below) |
| input | Context.gets(ctx, prompt), Context.read_secret(ctx, prompt) |
| piped stdin | Context.read(ctx) (drain to EOF), Context.read_chunk(ctx) |
| terminal | Context.terminal(ctx) → %{cols, rows, color, *_tty} |
| files | Context.file_read/file_write/file_append/file_delete |
| dirs | Context.dir_list/dir_create/dir_delete |
| env | Context.env(ctx, "NAME") |
| browser | Context.browser_launch(ctx, url) |
| raw input | Context.raw_input(ctx, fun), Context.read_key(ctx) — REPL/TUI |
| exit code | return an integer from run/1 (or Context.exit(ctx, n)) |
Files / env / browser are requests the client enforces against its per-app entitlement policy — your server can't touch the user's machine unless they grant it.
Parsing args — pick any style
Terminalwire hands you raw argv (Context.args/1); parsing is pure, so use whatever
you like. All three below work unmodified.
Raw / stdlib. Pattern-match, or use stdlib OptionParser for flags:
{opts, args, _} = OptionParser.parse(Context.args(ctx), strict: [verbose: :boolean])
Optimus — subcommands, typed args, generated
--help. Use Optimus.parse, never Optimus.parse!: the bang version calls
System.halt on --help/errors, which would take down your server. Handle the
result and render it yourself:
case Optimus.parse(spec(), Context.args(ctx)) do
{:ok, [:deploy], %{args: %{env: env}}} -> deploy(ctx, env)
:help -> Context.puts(ctx, Optimus.help(spec())); 0
{:error, errs} -> Enum.each(errs, &Context.warn(ctx, &1)); 1
end
Owl — rich UI (tables, color, prompts, spinners,
progress). It writes to the group leader, so it streams over the wire for free —
and it's width-aware: it asks the group leader for :io.columns, which Terminalwire
answers with the client's terminal width.
Owl.IO.puts(Owl.Table.new(rows)) # a table, rendered on the client
Owl.IO.puts(Owl.Data.tag("done ✓", :green)) # color
Owl.Spinner.run(fn -> deploy() end, labels: [...]) # live spinner
The standard "nice Elixir CLI" stack — Optimus to parse + Owl to render — works as-is over the wire.
Two rules (both about output, not parsing)
- Never
System.halt(orOptimus.parse!, or escript-style exits). Your handler runs inside the server; halting kills the BEAM. Return an exit code fromrun/1instead. - stdout is the group leader; stderr is not.
IO.puts/Owl.*/Context.putsreach the client (stdout). BareIO.puts(:stderr, …)goes to the server's console — useContext.warn/2for the client's stderr. (This is just Erlang's IO model::stderris a separate device from the group leader, not a Terminalwire quirk.)
Runnable examples
examples/self_describing.exs— a tiny CLI that streams its own source (rawContext.args+IO.ANSI).examples/owl_cli.exs— the full stack: Optimus subcommands + Owl tables/color/spinner.
Run either, then point a launcher at it:
elixir examples/owl_cli.exs
printf '#!/usr/bin/env terminalwire-exec\nurl: "ws://localhost:8081/terminal"\n' > app && chmod +x app
./app apps # an Owl table, streamed from Elixir to your terminal
Architecture
| layer | module |
|---|---|
| sans-IO protocol core | Terminalwire.Protocol, Terminalwire.Codec, Terminalwire.Negotiator, Terminalwire.Frames |
| sans-IO server state machine | Terminalwire.Server.Connection |
| process that drives it | Terminalwire.Server.Session |
| CLI-facing API | Terminalwire.Server.Context |
| command router (Thor-style) | Terminalwire.CLI |
| WebSocket adapter | Terminalwire.WebSock |
The protocol core mirrors the Ruby Terminalwire server and the Go client,
and is validated against the same language-neutral conformance corpus in
terminalwire/protocol — run mix test with TERMINALWIRE_CORPUS pointed at it.
That corpus is the cross-implementation contract: pass it and this server
interoperates on the wire with the client and every other server.
License
Apache-2.0 (source-available — safe to install on your own servers).