the future
of thinking
Our Rate

An AI agent that works as fast as your mind, right in your terminal, no fuss, no limits.

TerminalAgent
↗
3Supported
Platforms
5Built-in
Tools

RUNS WITH:

npm npx bun macOS Linux Windows

We make coding faster, smarter, and accessible for everyone — from curious beginners to power users.

AI × Terminal: one
future, one force

Power your workflow with an agent that reads your code, edits your files and runs your commands — on macOS, Linux and Windows. Start it with npx, npm or bun, from first idea to shipped release.

01 — How it works

From prompt to pull request.

01

Ask

Describe the task in plain words — a bug, a feature, a refactor. No setup, no config files.

❯ the login test is failing, fix it
02

Read

SONNA explores the project first: lists folders, opens the files that matter, runs the failing test.

◆ list_dir .
◆ read_file src/auth.js
◆ $ npm test
03

Act

Exact, unique edits for small changes. Whole files for new ones. Every write waits for your yes.

◆ edit_file src/auth.js
? edit src/auth.js? [y/N] y
04

Verify

It runs your tests again and reports back in plain words — what changed, and why it now works.

✓ 24 passing (412ms)
Fixed: expiry now uses
Date.now()/1000.
02 — Capabilities

Six tools. Zero noise.

Everything an engineer does at a keyboard, with nothing it doesn't need.

◆

Reads before it acts

Explores your project with list and read tools before it changes a single line.

list_dir · read_file
◆

Precise edits

Replaces one exact, unique snippet — or writes a whole new file when that is the job.

edit_file · write_file
◆

Runs your shell

Tests, builds, git. cmd.exe on Windows, /bin/sh on macOS and Linux.

run_command
◆

Asks first

Every write and every command waits for your yes. Pass --yes only when you trust the task.

[y/N]
◆

Thinks deeply

Claude Opus 5.5 with adaptive thinking — it reasons as hard as the problem needs.

claude-opus-5-5
◆

Stays in the workspace

Paths are checked against the project root. Nothing outside it can be read or written.

sandboxed paths
03 — Live session

Scroll. Watch it work.

Every line on the right is a real SONNA session — scroll to play it.

  ~/shop-api — sonna
$ npx sonna-agent
SONNA AGENT · claude-opus-5-5 · ~/shop-api
 
❯ add rate limiting to the /login route
 
◆ list_dir src
◆ read_file src/routes/auth.js
◆ read_file package.json
No limiter installed. I'll add a small in-memory one.
◆ write_file src/middleware/rateLimit.js
? write src/middleware/rateLimit.js (612 chars)? y
◆ edit_file src/routes/auth.js
? edit src/routes/auth.js? y
◆ $ npm test
✓ 31 passing (688ms)
 
Done: /login now allows 5 attempts per minute per IP.
❯
0Platforms
macOS · Linux · Windows
0Runtimes
npm · npx · bun
0Built-in
tools
0Dependency
no build step
04 — Install

One line. Any terminal.

Node.js 18+ or Bun. Set ANTHROPIC_API_KEY and begin.

macOS

$ export ANTHROPIC_API_KEY=sk-…
$ npx sonna-agent

Linux

$ export ANTHROPIC_API_KEY=sk-…
$ bunx sonna-agent

Windows

> $env:ANTHROPIC_API_KEY="sk-…"
> npm i -g sonna-agent; sonna
05 — Questions

Asked & answered.

The agent is MIT-licensed and free. It calls the Claude API with your own key, so you pay only for the API usage of your sessions.

Claude Opus 5.5 by default, with adaptive thinking. Pick another with --model or the SONNA_MODEL environment variable.

No. Every file write and every shell command waits for your y. Only --yes turns that off.

Yes. Commands run in cmd.exe on Windows and /bin/sh on macOS and Linux, and SONNA tells the model which shell it is in.

Pass the task as an argument — sonna "update the changelog" — and it runs once and exits with a status code.

06 — Begin

Start thinking.

One command. Your terminal. A calm agent that gets it done.