A coding agent for the terminal, in a 3 MB binary.

claude-jr, or Claude Junior, does what Claude Code does: it reads your code, edits files and runs commands, asking before anything changes. It is written in Rust with no runtime, so it starts in 5 ms and sits in 2 MB of memory.

curl -fsSL https://claudejr.iwhitebird.com/install | sh

Linux and macOS, on x86 and ARM; on Windows, inside WSL. Then run claudejr and paste an Anthropic API key.

claude-jr 0.1.3
claude-opus-5-5 · ~/proj/cli
/help for commands · shift+tab to change mode> add a --verbose flag to the cli

● Bash(grep -n "struct Args" -A12 src/cli.rs)
  ⎿ 9:pub struct Args {  … +11 lines

● Edit(src/cli.rs)
    13 + /// Print debug output
    14 + #[arg(short, long)]
    15 + pub verbose: bool,
  ⎿ Edited src/cli.rs (1 replacement)

● Bash(cargo build)
  ⎿ Finished `dev` profile in 2.1s

● Added --verbose to Args. The build passes.> ask anything, or /help

Small enough not to notice

There is no Node or Python to install and nothing running in the background. A whole working session, including the commands it runs, fits in a 16 MB container on a twentieth of a CPU core.

3.2 MBone static binary

2 MBmemory when idle

5 msto start

0 msCPU while it waits

	claude-jr	Claude Code	pi
Install size	3.2 MB	324.5 MB	109.0 MB
Memory, idle	2.0 MB	135.0 MB	141.1 MB
Memory, working	2.5 MB	270.2 MB	137.5 MB
Startup	5 ms	167 ms	~260 ms
Its own CPU, 20 tool calls	30 ms	1,150 ms	580 ms

All three run against the same fake API, so the model's time isn't counted; how it's measured.

And it gets the work done

SWE-bench Verified gives an agent real GitHub issues from projects like Django, sympy and scikit-learn, and checks its fix with the projects' own hidden tests.

88%

of a random sample of 100 tasks resolved with Claude Sonnet 5.5, one attempt each, at about five cents a task. With Claude Opus 5.5, 45 of 50 (90%).

What's in it

Subagents: Side tasks go to agents that work in parallel and report back.
Permission modes: From asking about everything to an auto mode where a small model reviews each command. shift tab switches.
A sandbox: With --sandbox on Linux it can only write to your project; keys, sockets and the rest of your home stay out of reach.
Rewind and resume: Go back to before any message with its file changes undone. Every session is saved as it goes.
Skills and MCP: Reads your .claude/skills and talks to MCP servers, local or remote.
Images and the web: Drag in a screenshot, and it can search and fetch pages when the task needs them.
A stash: ctrl s puts a half-written message aside while you send another.
Scripting: claudejr -p "…" runs one prompt, with JSON output and a spending limit for CI.

Using it

`claudejr`	start in this folder
`claudejr -c`	continue the last conversation
`claudejr -p "explain main.rs"`	one prompt, then exit
`claudejr --sandbox`	only write inside this folder
`claudejr --update`	get the newest release
`/`	every command, inside it