Beledarian's LM Studio Tools

English | Deutsch | 简体中文 | 繁體中文

This project is a plugin for LM Studio that provides a rich set of tools to a large language model. It acts as a bridge between the LLM and your local environment, enabling autonomous coding, research, and file management.

Important

LM Studio does NOT support automatic updates. If you encounter issues, please try manually updating first by removing the current version and re-downloading from the plugin website. LM Studio may show an "already installed" tooltip, even if your version is outdated.

Key Features

File System Mastery

• Full Control: Create, read, update, delete, move, and copy files.
• Safe & Secure: All operations are sandboxed to your workspace directory to prevent path traversal attacks.
• Smart Updates: Use replace_text_in_file to make surgical edits instead of rewriting large files.
• Batch Processing: save_file supports creating multiple files in one go.
• Cleanup: Use delete_files_by_pattern to wipe temporary files instantly.

Encountering issues? Feel free to submit them on GitHub.

Find this project helpful? Consider giving it a ⭐ on GitHub or contributing! Thank you for using the toolbox.

Recent Updates (v1.3.1)

• 🌍 Full Internationalization (i18n): Added complete UI and runtime support for English, German, Simplified Chinese, and Traditional Chinese.
• 🌐 Dual-Layer Translation: Supports both "Config UI" (static) and "Agent Messages" (dynamic runtime) languages.
• 🔄 UI Language Override: Added a field to manually force the Config UI into a specific locale for testing, bypassing OS-level detection on the next plugin restart.

Older Updates (v1.3.0 & earlier)

v1.3.0

• 🐙 Native GitHub CLI Tools: Added gh_auth, gh_create_issue, gh_list_issues, gh_view_comments, gh_create_pr, gh_list_prs, gh_view_pr_diff, and gh_push for secure, structured GitHub interactions without generic shell commands
• 🌿 Enhanced Git Workflow: Added git_add and git_checkout tools to complete the native git toolchain (status, diff, log, commit, add, checkout)
• ⚙️ New Sidebar Toggle: Added "Allow GitHub CLI Tools" setting to enable/disable gh_* tools independently
• 🛡️ Dependency Guards: All CLI tools now verify installation before execution and provide clear prompts if missing

v1.2.0

• 🛡️ Sub-Agent Tool Validation: Added early parameter validation with clear error messages to prevent silent failures when sub-agents use wrong parameter names or absolute paths outside workspace
• 🧪 Regression Tests: Added 14 new tests for tool call validation logic (total: 51 tests)
• 💬 Better Error Feedback: Sub-agents now receive TOOL_VALIDATION_ERROR messages with helpful hints about correct parameter names vs common mistakes

v1.1.1 (2026-04-08)

Browser Reliability and Navigation Context

• Fixed: Browser action clicks now include a DOM-level fallback when Puppeteer reports "Node is either not clickable or not an Element"
• Improved: Browser click actions now retry native click after ~300ms before falling back
• Added: browser_session_open returns full page text by default (include_page_text defaults to true)
• Refined: Multi-step routing guidance now prioritizes browser_session_open -> browser_session_control -> browser_session_close

v1.1.0 (2026-04-08)

Sub-Agent Compatibility Improvements

• Fixed: Gemma 4 and other models using {"tool": "...", "parameters": {...}} format now work correctly with consult_secondary_agent
• Added: Advanced browser navigation (browser_session_open, browser_session_control, browser_session_close) including in-page fuzzy find and URL-change notices
• Added: Structured sub-agent handoff message support (handoff_message) for relay/summary workflows
• Added: Enable Sub-Agent Debug Logging toggle in plugin settings
• Added: Support for direct {file_name, content} JSON format from some models

v1.0.x Previous Updates

• Smart Context Injection: subagent_docs.md is automatically loaded into the context
• Enhanced Reporting: Fixed file path reporting in consult_secondary_agent
• Project Tracking: Sub-agents enforce creation of beledarian_info.md
• Strict Naming: Improved instructions for correct file extensions

Autonomous Agents

• Secondary Agent: Delegate complex tasks (coding, summarization) to a second local model/server with support for the main model already used and loaded by LM Studio!
• Auto-Save: When the sub-agent generates code, the system automatically detects and saves it to your disk. No more copy-pasting!
• Auto-Debug: (Optional) Triggers a "Reviewer" agent to analyze generated code and fix bugs automatically before returning the result.
• Structured Handoff: Sub-agents can return a dedicated handoff_message for the main agent to relay findings/research.
• Project Context: Agents can read beledarian_info.md to understand your project's history.

Code Execution

• Sandboxed: Run JavaScript (Deno) and Python code.
• Terminal: Execute shell commands or open real terminal windows for interactive tasks.

Warning

Enabling shell or terminal execution allows the model to run arbitrary commands on your system. If enabled, the model may be able to modify files and escape the sandbox environment.

Web & RAG

• Research: Search DuckDuckGo, Wikipedia, or fetch raw web content.
• Advanced Browser Navigation: Persistent browser_session_open/control/close flow for multi-step browsing and automation.
• Web RAG: Chat with website content.
• Local RAG: Semantic search over your workspace files (rag_local_files).

Requirements

• Node.js (v18+)
• LM Studio (v0.3.0+)

💡 Tip: Need persistent long-term memory for your agent? Check out my other project: Local Memory MCP – A privacy-first memory server with knowledge graph support.

Installation

The plugin can be installed using the following link:

https://lmstudio.ai/beledarian/beledarians-lm-studio-tools

Alternatively, you can install it manually for development purposes.

Development

If you want to contribute to the development of this plugin, you can follow these steps:

1. Clone the repository:

   git clone https://github.com/Beledarian/Beledarians_LM_Studio_Toolbox.git
   cd Beledarians_LM_Studio_Toolbox

2. Install dependencies:

   npm install

3. Run in development mode: From within the project directory, run the following command:

   lms dev

   This will start the plugin in development mode. LM Studio should automatically pick it up. Any changes you make to the source code will cause the plugin to automatically reload.

Configuration

Access these settings in the LM Studio "Plugins" tab:

• Enable Secondary Agent: Unlock the power of sub-agents.
• Sub-Agent Profiles: Custom prompts for "Coder", "Reviewer", etc.
• Auto-Debug Mode: Automatically review sub-agent code.
• Sub-Agent Debug Logging: Toggle detailed sub-agent parsing logs for troubleshooting.
• Sub-Agent Auto-Save: Toggle automatic file saving (Default: On).
• Show Full Code Output: Toggle whether to display the full code in chat or hide it for brevity (files are still saved).
• Default Workspace Path: Set the startup workspace directory used by the plugin.
• Safety: Enable/Disable "Allow Code Execution" for Python/JS/Shell.
• Browser Safety: Browser automation for sub-agents requires all three toggles: Allow Browser Control + Sub-Agent: Allow Web Search + Sub-Agent: Allow Browser Control.

Available Tools

File System

• list_directory, change_directory, make_directory
• read_file, save_file (supports batch), delete_path
• replace_text_in_file: Precision editing.
• delete_files_by_pattern: Regex-based cleanup.
• move_file, copy_file, find_files, get_file_metadata
• fuzzy_find_local_files: Levenshtein-based fuzzy file path/name search.

Agent

• consult_secondary_agent: The powerhouse tool. Delegates tasks, handles file creation, and manages sub-agent loops.

Web

• web_search (DuckDuckGo API + HTML fetch/browser fallback), wikipedia_search
• fetch_web_content, rag_web_content
• browser_session_open, browser_session_control, browser_session_close (persistent page automation; preferred for multi-step navigation, with deduped page-text output unless full_read=true)
• browser_open_page (stateless one-shot Puppeteer read)
• Workaround tip: if selector-based click fails, use an evaluate action to click by text, then call with full_read=true.

{
  "actions": [
    {
      "type": "evaluate",
      "script": "const link=[...document.querySelectorAll('a')].find(a=>a.textContent?.includes('LICENSE')); if(link) link.click();"
    }
  ],
  "full_read": true
}

Execution

• run_javascript, run_python
• execute_command (Background), run_in_terminal (Interactive)

Utils

• rag_local_files: Search your code.
• save_memory: Long-term memory.
• get_system_info, read_clipboard, write_clipboard

Developer Guide

See CODE_OVERVIEW.md for architectural details.