An open-source, model-agnostic AI coding CLI. Lemonade is a Claude Code–style agentic terminal assistant that runs against any model on OpenRouter — pick whichever model you have credits or access for. Point it at an API key, describe what you want, and it plans, writes files, runs commands, and iterates until your project runs. All wrapped in a warm citrus terminal UI.
# Run it without installing
npx lemonade-code
# …or install globally (the command is `lemonade`)
npm install -g lemonade-code
lemonadeRequires Node.js 20+.
The first time you run lemonade, it skips straight to a short setup instead
of the normal welcome screen:
- Your name — what Lemonade should call you (defaults to your OS username).
- Your OpenRouter API key — paste it in, or press Enter on an empty line to open https://openrouter.ai/keys in your browser, grab a key, and paste it back. Lemonade validates the key live before saving it.
That's it — no separate setup command. Your name and key are saved to
~/.lemonade/config.json (the key is masked as sk-or-...ab12 anywhere it's
shown, and never printed in full).
Change them any time:
/settings— update your name, rotate your API key, set the default model, toggle auto-approve or animations, and view your OpenRouter credit balance./logout— clear the saved key and return to the login screen on next launch.
You can also inject a key without saving it (handy for CI):
LEMONADE_API_KEY=sk-or-... lemonade --yes "run the test suite"lemonade
# log in once, then just talk to it:
> build a REST API for a bookstore with Express and SQLite
> add a /books/:id endpoint with tests
> /model # switch to a cheaper or free model
> /folder ./api # cd into a subproject
> npm test # ask it to run and fix failing testsLemonade will read and search your files, propose writes and edits (showing you a diff), run commands (asking before anything that isn't read-only), and narrate each step — just like a pair programmer.
/help is the source of truth and lists everything dynamically. The core set:
| Command | Description |
|---|---|
/help |
List all available commands. |
/settings |
Change username, rotate API key, set default model, toggle auto-approve/animations, view credits. |
/model |
List and switch the active OpenRouter model (free vs. paid, context length, pricing). |
/init |
Scaffold a new project in the current directory. |
/folder (/cd) |
Change the directory Lemonade operates in (creates it if missing). |
/agents |
Spin up a named subagent for a scoped task, e.g. /agents reviewer check src/ for bugs. |
/history (/resume) |
Show or resume past sessions in this project. |
/clear |
Clear the current conversation context (files untouched). |
/undo |
Revert the last file change Lemonade made. |
/logout |
Clear the saved API key. |
Keyboard: Esc or Ctrl+B interrupts a running task · Ctrl+C twice exits.
| Flag | Meaning |
|---|---|
[directory] |
Directory to operate in (defaults to the current directory). |
-y, --yes |
Auto-approve non-destructive tool calls (for CI/trusted use). |
--dangerously-skip-confirmations |
Skip all confirmations, including destructive commands. Opt-in and loud on purpose. |
--no-animations |
Disable startup/streaming animations (nice over slow SSH). |
-v, --version |
Print the version. |
Even with --yes, genuinely destructive commands (rm -rf, force pushes, disk
formatting, sudo, pipe-to-shell, …) still require confirmation — only
--dangerously-skip-confirmations bypasses them.
Lemonade works with any model OpenRouter exposes — it never hardcodes a model
roster. Browse the full, always-current list at
https://openrouter.ai/models, or run /model to pick one interactively (it
labels free vs. paid, shows context length and pricing, and lets you filter, e.g.
type free or claude).
Free models (the :free suffix) are labeled [free]. They have tighter rate
limits; when you hit one, Lemonade retries with backoff and, if it still fails,
shows a clear message pointing you to switch models or check your dashboard.
OpenRouter is the default, but Lemonade talks the OpenAI chat-completions
protocol, so it also works with self-hosted endpoints like Ollama or
LM Studio. Set a baseUrl in ~/.lemonade/config.json (or the
LEMONADE_BASE_URL env var):
{ "baseUrl": "http://localhost:11434/v1", "model": "llama3" }Lemonade runs a standard agent loop: it sends your message plus tool definitions to the model, executes any tool calls locally, feeds results back, and repeats until the model returns a final answer.
Tools: read_file, write_file, edit_file (unique-match, diff-previewed),
list_directory, run_command (risk-classified), and search_files
(ripgrep-style). All file and shell operations are sandboxed to the working
directory by default. Every change is logged so /undo and /history work.
git clone https://github.com/lemonade-cli
cd lemonade
npm install
npm run build # compile TypeScript to dist/
npm test # run the vitest suite
npm run dev # tsc --watch
node dist/cli.js # run your local buildThe repo ships zero API keys — .gitignore excludes .lemonade/ and .env.
Never commit a real key.
Contributions are welcome! See CONTRIBUTING.md and our
LICENSE© Lemonade.