mini-claude-code

AI coding assistant with direct filesystem access. Learning project.

---

Quick Start

bunx github:thatbeautifuldream/mini-claude-code

npx github:thatbeautifuldream/mini-claude-code

pnpm dlx github:thatbeautifuldream/mini-claude-code

yarn dlx github:thatbeautifuldream/mini-claude-code

export OPENAI_API_KEY=your-key-here

---

Local Dev

bun install

bun run dev

---

Screen.Recording.2026-02-26.at.10.51.10.PM.mov

---

How It Works

Uses ToolLoopAgent from ai SDK:

const agent = new ToolLoopAgent({
  model: openai("gpt-4o"),
  tools: bashToolkit.tools,
  instructions: INSTRUCTIONS,
  stopWhen: stepCountIs(50),
});

Each turn streams steps via onStepFinish:

agent.stream({
  messages: trimmed,
  onStepFinish(stepResult) {
    // Captures tool calls, results, token usage
  },
});

Bash tool wraps just-bash with safety hooks:

createBashTool({
  sandbox: customBash,
  onBeforeBashCall: blockDangerousCommands,
  onAfterBashCall: logFailures,
});

Direct filesystem via ReadWriteFs:

new ReadWriteFs({ root: rootPath });

Custom commands:

defineCommand("explain", readFile);
defineCommand("lint", suggestLinter);
defineCommand("deps", showDeps);

---

Commands

Command	What
/help	Show commands
/clear	Clear history
/git	Git status + commits
/exit	Quit

---

Safety

Blocks DANGEROUS_COMMANDS via onBeforeBashCall:

• rm -rf /, rm -rf *, rm -rf .
• dd if=, mkfs, fdisk
• git push --force, git reset --hard

---

Author

Milind Mishra

---

License

MIT