LemonCode 🍋

A friendlier window around Claude Code: type what you want and press Enter, your own templates, a live token meter, a one-click COMPLETE button, and Citron, a pixel lemon who shows what Claude is doing.

Requirements

Windows 10 or 11 (the app uses claude.exe, PowerShell and a Desktop shortcut, so nothing else is tested), plus Node.js for npm install and Git for the project folders.

Start

git clone https://github.com/j99-droid/lemoncode.git
cd lemoncode
git config core.hooksPath .githooks
npm install
npm start

Or double-click LemonCode on the Desktop (npm run shortcut makes it).

LemonCode runs Anthropic's Claude Code CLI in a window, so you need a Claude plan or account. The first start opens Getting started on the home page. Pick one of three setups: Claude account (everything on Claude), Claude + local helpers (Claude as usual, helpers on a model here), or Local only (every session on a model here, no Claude account; see below). It checks Claude Code and Ollama and offers an Install button for each one that's missing. Claude Code's button runs Anthropic's official installer (irm https://claude.ai/install.ps1 | iex), and the first session of a Claude setup asks you to sign in with your Claude account. Reopen the panel any time from ⋯ → Getting started.

Citron's reactions come from a small hook server that LemonCode starts on 127.0.0.1 (a random local port). Only sessions LemonCode starts are sent to it, but it has no password, so other programs on your PC could post to it.

Local helpers (optional)

A helper can run on a model on your PC instead of Claude. It needs Ollama (0.14 or newer) and the GPT-OSS 20B model. Getting started installs Ollama with winget and downloads the model for you, or do it by hand:

ollama pull gpt-oss:20b
ollama create lemon-gpt-oss-20b -f tools/local-model/Modelfile

Then pick Local: GPT-OSS 20B as a helper's model. Only sessions with a local helper use the local router (src/router.js): it sends that helper's calls to Ollama and everything else to Anthropic, so your Claude sign-in is unchanged. The main Claude is always Claude. Ollama must be running when the helper is used (it runs in the tray after install). The model is about 13 GB and fits fully in 16 GB of GPU memory.

Local helpers are told to work one job at a time. Ollama answers one call at a time with one prompt cache, so two local helpers side by side queue behind each other and keep re-reading their 15-30K token prompts: in testing on an RTX 4080 Super, calls took 30-90 s instead of about 2 s.

Local only (the setup panel) runs every session on that model instead: no Claude account is used, and calls go to Ollama only. Claude Code is still needed as the program that runs the session. The model is smaller and weaker than Claude, and its context window is 64K, so long sessions lose their oldest messages.

What's on screen

• Type box (home page): type what you want, press Enter. With Blank picked, LemonCode makes a folder in the projects folder (default ~\Lemon Projects) named from your words, with git, AGENTS.md, CLAUDE.md (@AGENTS.md), .gitignore and the secrets pre-commit hook, marks it trusted in ~/.claude.json (backup in ~/.claude.json.lemon-backup) and starts Claude with your words as the first prompt.
• Helpers (under the type box): tick which helper agents join the project. They're passed to Claude Code as subagents (--agents), each with its own role, model and effort, plus a short system-prompt note telling the main Claude to hand matching work to them. A resumed session gets the same helpers back.
• Templates (toggle under the helpers, closed each time the app opens): Blank is built in; ＋ New template saves a name, icon, folder (new each time, or always the same), the helpers it brings and a starting prompt.
• Tabs: ⌂ New project (home) first, then one tab per running session (dot = Citron's state for it, plus how many helpers are working). ⋯ opens Agents (make, edit, delete helpers; "Add starter team"), Recent (all Claude Code sessions to reopen or archive) and Archived projects.
• Chat view (each session opens in it): what you and Claude say, as chat bubbles, plus each helper as its own member of a group chat: its icon, when it starts and finishes, what it is doing (reading, editing, running) and what it says, live as it works. Claude's own tool calls and their output are left out. Type in the box at the bottom: Enter sends, Shift+Enter is a new line, Esc stops Claude, and pasting an image adds it to your next message. Terminal in the session header switches to Claude's real terminal and back.
• Question cards: when Claude asks you something, the chat shows a card with each question's options (tick boxes for pick-many questions, and "Type something" for your own answer). Option previews show as Markdown, code, or live HTML in a sandboxed frame with no network. Send answer types your choices into Claude's dialog. The terminal still opens by itself for approvals, sign-in and "trust this folder?", and goes back to chat after.
• Model / Effort / Mode: for the main Claude, applied to sessions started after you change them. Model and Effort each have a slider (lists in src/power.js). The arm beside them grows with both: a weak arm for Haiku at low effort, a huge arm with fire for Fable at max.
• 5h and Week bars: your Claude plan's 5-hour and weekly limits as a percent, with reset times, as Claude Code reports them to its status line (Pro/Max plans, after the first reply; the last values are kept across restarts). Click a bar for that window's tokens, sessions, cache reads and a per-model breakdown. Those token totals come from this PC's transcripts, so they won't add up to the percent, which also counts claude.ai and other computers. Local-model calls are shown separately and left out. Hidden in Local only.
• COMPLETE: asks "Is that all?", offers a commit if there are changes (never pushes), closes the session and moves it to Archived projects. If git has no name/email yet it asks once and saves them globally.
• Citron: reacts to hooks from sessions started here: thinking, typing, coding (with what it's doing), focused (long tool runs), happy (finished), chill (quiet a minute), needs-you (approval or question). When the window isn't in front, approvals also pop a Windows notification.

Files

File	What it does
src/main.js	Window, Claude terminals (node-pty), chat feeds, hook listener, git, project creation
src/usage.js	Reads transcripts for token totals per limit window and the session list
test/usage.test.js	Tests for the usage totals (npm test)
src/statusline.js	Claude Code status line command: passes the 5h and weekly limits to the app
test/statusline.test.js	Tests for reading the limits (npm test)
src/chat.js	Follows one session's transcript and its helpers' transcripts for the chat view
test/chat.test.js	Tests for which transcript lines the chat view shows (npm test)
src/activity.js	Short labels for tool calls ("Editing notes.md"), for Citron and helper bubbles
src/questions.js	Turns a question card's answers into the keys Claude's question dialog reads
test/questions.test.js	Tests for those keys (npm test)
src/router.js	Local model router: sends a local helper's calls to Ollama, the rest to Anthropic
test/router.test.js	Tests for where the router sends each call (npm test)
tools/local-model/Modelfile	Makes the lemon-gpt-oss-20b model with a 64K context (see Local helpers)
src/setup.js	Getting started checks: Claude Code, Ollama and the local model, and the install and pull commands
test/setup.test.js	Tests for what counts as ready and how Ollama's answer is read (npm test)
src/hook.js	Hook command: forwards Claude's hook JSON to the app
src/templates.js	Folder naming, prompt building, AGENTS.md skeleton
src/power.js	The model and effort slider lists, in slider order, and the defaults
src/renderer/arm.js	Draws the flexed arm on the Power slider (SVG in code, fire at the top step)
AGENTS.md / CLAUDE.md	Rules for AI agents working on this repo (CLAUDE.md just loads AGENTS.md)
.githooks/pre-commit	Blocks committing secrets, databases and logs (enabled by the core.hooksPath line above)
LICENSE	MIT
src/preload.js	Safe bridge between page and main process
src/renderer/	The page: index.html, styles.css, app.js, citron.js (the sprite)
src/renderer/chatview.js	Draws chat bubbles, helper bubbles and question cards (Markdown, previews)
src/renderer/preview-frame.html	Holds one HTML preview from a question card, sandboxed with no network
preview/citron-preview.html	Sprite sheet of every pose (serve the folder over http to view)
tools/make-icon.js	Redraws build/icon.png / icon.ico from Citron
tools/make-shortcut.ps1	Makes the Desktop shortcut
templates/githooks/pre-commit	Secrets hook copied into new projects

Settings

Stored in %APPDATA%\lemoncode\config.json: daily budget, archived sessions, model, effort, mode, projectsRoot (default ~\Lemon Projects), your templates and agents, which helpers each session had, and an optional claudePath.

License

MIT. See LICENSE.