quarkcode

Minimal CLI agent with a pretty TUI, 855 lines of code

Overview

This probably won't replace your current agent, but it's a fun experiment in building a minimal agent with a TUI. It's functional and somewhat polished, but not fully featured.

Inspired by mini-swe-agent and Claude Code, the goal was to build an equivalent agent in under 1k lines (mini-swe-agent is ~1.7k lines for the CLI + TUI).

Fun fact: Claude Code helped write the initial non-interactive version, then I switched to quarkcode itself to build the TUI and polish the features.

Why "quark"? Among the most fundamental particles we know. And just a unique name.

Details

• Interactive TUI similar to Claude Code
• One-shot sessions (will not remember previous prompts within your session)
• Single tool: Bash execution
• Auto-accept mode (Shift+Tab)
• Potentially resolves ~67% of GitHub issues in SWE-bench verified (unverified claim)
• Supports Anthropic and OpenAI models
• Written in TypeScript
• Minimal dependencies
  • AI SDK
  • Ink (TUI)
  • yargs (CLI)

Quickstart

⚠️ Warning: Be careful with auto-accept mode. This tools gives an LLM shell access.

# Set ANTHROPIC_API_KEY or OPENAI_API_KEY in your environment
export ANTHROPIC_API_KEY="your_key_here"

# Install globally
npm install -g quarkcode@latest
quark                    # Interactive mode
quark -t "Fix the bug"   # Direct task

# Or run directly without installing
npx -y quarkcode@latest  # Interactive mode
npx -y quarkcode@latest -t "Fix the bug in main.py"

# Or via bunx
bunx quarkcode  # Interactive mode
bunx quarkcode -t "Fix the bug in main.py"

TUI Shortcuts

• Ctrl+C - Exit
• Shift+Tab - Toggle auto-accept mode (yolo mode with no safeguards)
• Command confirmation: Y/Enter = approve, N/Escape = reject

CLI Flags

• -t, --task - Task to execute in non-interactive mode
• -m, --model - LLM model
• -l, --token-limit - Max tokens
• -s, --step-limit - Max steps
• --tui - Launch TUI
• -h, --help - Show help

Command Aliases

Both quark and quarkcode work identically:

• quark -t "task" (short)
• quarkcode -t "task" (descriptive)

Models

• claude-sonnet-4-20250514 (default)
• gpt-5
• any other OpenAI or Anthropic model (untested)

Quickstart in a docker container

Work on any project by mounting it into /workspace:

# 1. Build image (one time)
git clone <repo> && cd quarkcode && docker build -t quarkcode .

# 2. Configure API key
echo "ANTHROPIC_API_KEY=your_key_here" > .env

# 3. Set your project path
export PROJECT_DIR="/path/to/your/project"

# 4. Run tasks
docker run --rm -it -v $PROJECT_DIR:/workspace --env-file .env -w /workspace quarkcode -t "Fix the bug in main.py"

# 5. Interactive mode
docker run --rm -it -v $PROJECT_DIR:/workspace --env-file .env -w /workspace quarkcode

Local Development

bun run agent will run in docker.

# Interactive TUI in docker (default when no task provided)
bun run agent

# Direct task execution
bun run agent -t "Create a simple React component"

# Build + run
bun run agent:b