LangTARS — Native Claw-Like Plugin

简体中文  •   繁體中文  •   日本語  •   GitHub

⚠️ Note: Features are still under active development. If you encounter any bugs, please submit an issue.

---

What is LangTARS?

LangTARS is a native Claw-like plugin inspired by Nanobot's ReAct philosophy, designed to bring the OpenClaw experience to LangBot users. It enables you to control your Mac through IM messages using autonomous AI task planning. Like TARS from Interstellar, it works faithfully for you.

Like OpenClaw, LangTARS allows AI assistants to execute real actions on your Mac—but with the simplicity and elegance of a LangBot plugin.

Why LangTARS?

OpenClaw is an impressive project with a great vision. However, running such complex software with access to your digital life requires trust in systems you may not fully understand.

LangTARS takes a different approach:

• Native LangBot integration — Runs directly within LangBot using the Nanobot kernel
• Lightweight — Minimal codebase you can understand and audit
• Autonomous planning — Uses ReAct loop for intelligent task execution
• Safety-first — Built-in command restrictions, workspace isolation, and dangerous command blocking

Quick Start

1. Install LangTARS through LangBot's plugin system
2. Configure your preferred LLM model for task planning
3. Start controlling your Mac via IM messages!

Main Command

!tars auto — Autonomous Task Planning

This is the primary command that makes LangTARS special. Simply describe what you want to do, and the AI will autonomously plan and execute the task using available tools.

• !tars auto Open Safari, visit langbot.app, scrape elements and tell me

• !tars auto Create a new note with title and content "hello"

• !tars auto Help me organize the files on my desktop

• !tars auto Cook me some dishes.

Browser Control

LangTARS supports three browser control methods:

Command Example	Browser	Description
!tars auto Visit github.com	Playwright (Chromium)	Default, no extra permissions needed
!tars auto Open Safari and visit github	Safari Browser	Uses real Safari, requires AppleScript permission
!tars auto Open Chrome and visit github	Chrome Browser	Uses real Chrome, requires AppleScript permission

The AI will:

1. Understand your request
2. Plan the necessary steps
3. Execute them one by one using shell commands, file operations, app control, etc.
4. Report back with results

You can stop a running task at any time:

!tars stop

Check task status:

!tars status

Testing Commands

These commands are available for testing and direct control:

Command	Description
!tars auto <task>	Autonomous task planning (AI-powered)
!tars shell <command>	Execute a shell command
!tars ps [filter]	List running processes
!tars kill <pid|name>	Kill a process
!tars ls [path]	List directory contents
!tars cat <path>	Read file content
!tars open <app|url>	Open an application or URL
!tars close <app>	Close an application
!tars apps [limit]	List running applications
!tars info	Show system information

Configuration

Configure LangTARS through LangBot's settings:

Option	Description	Default
allowed_users	User IDs allowed to control this Mac	[]
command_whitelist	Allowed shell commands (empty = all with restrictions)	[]
workspace_path	Working directory for file operations	~/.langtars
enable_shell	Enable shell command execution	true
enable_process	Enable process management	true
enable_file	Enable file operations	true
enable_app	Enable app control	true
enable_applescript	Enable AppleScript execution	true
enable_browser	Enable browser automation (Playwright)	true
browser_type	Browser engine (chromium/firefox/webkit)	chromium
browser_headless	Run browser in headless mode	false
browser_timeout	Browser operation timeout (seconds)	30
planner_max_iterations	Max ReAct loop iterations	5
planner_model_uuid	LLM model for task planning	(first available)
planner_rate_limit_seconds	Rate limit between LLM calls	1
planner_auto_load_mcp	Auto-load MCP tools	true
planner_auto_load_skills	Auto-load skills from ~/.claude/skills	true
skills_path	Skills directory path	~/.claude/skills

Safety Features

• Dangerous command blocking — Commands like rm -rf / are blocked by default
• Workspace isolation — File operations restricted to configured workspace
• Command whitelist — Optionally restrict to specific commands
• User access control — Optionally limit to specific users

Architecture

IM Message --> LangBot --> PlannerTool (ReAct Loop) --> Tools --> Mac Actions

• PlannerTool — ReAct loop for autonomous task planning using LLM
• Tool Registry — Dynamic tool loading from MCP servers and plugins
• Built-in Tools — Shell, process, file, app control

License

MIT