high-perf-tools

A high-performance LM Studio plugin for Python and C++ development, filesystem operations, git, web research, and autonomous Agent Zero capabilities.

Features

Filesystem (9 tools)

Tool	Description
read_file	Read files with optional line range selection
write_file	Write or append to files, creating parent dirs
replace_in_file	Surgical find-and-replace (literal or regex)
delete_path	Delete files or directories
move_path	Move or rename files/directories
copy_path	Copy files within the workspace
list_directory	List files with metadata, filtering, recursive walk
create_directory	Create directories (with parents)
file_info	File metadata: size, line count, permissions
search_in_files	grep-like search across workspace files

Execution (4 tools)

Tool	Description
run_command	Execute arbitrary shell commands
run_python_code	Run a Python snippet (temp file)
run_python_file	Run an existing Python script
get_environment	Report Python version, PATH, installed tools

Python Dev (7 tools)

Tool	Description
analyze_python	AST analysis: imports, classes, functions, globals
format_python	Format with ruff or black
lint_python	Lint with ruff or pylint, optional auto-fix
run_tests	Run pytest with filtering and verbosity control
pip_install	Install Python packages
pip_list	List installed packages (supports --outdated)
type_check_python	Run mypy, optional --strict mode

C++ Dev (6 tools)

Tool	Description
compile_cpp	Compile with clang++/g++, sanitizers, custom flags
run_binary	Execute a compiled binary
compile_and_run	Compile and immediately run a snippet or file
format_cpp	Format with clang-format
analyze_cpp	Static analysis with clang-tidy
cmake_build	CMake configure + build
get_cpp_info	Compiler and tool version report

Research (2 tools)

Tool	Description
fetch_url	Fetch a URL and return plain text (HTML stripped)
search_web	DuckDuckGo search, returns titles + snippets

Git (8 tools)

Tool	Description
git_status	Working-tree status
git_diff	Working-tree or staged diff
git_log	Commit history with filtering
git_show	Show a specific commit
git_branches	List local and remote branches
git_stage	Stage files (git add)
git_commit	Create a commit (optionally stages files first)
git_init	Initialise a new repository

Agent Zero (8 tools)

Tool	Description
memory_save	Persist a fact or solution across conversations
memory_recall	Keyword-search persistent memory
memory_list	List all saved memories
memory_delete	Delete a memory by id or key
list_models	List models loaded in LM Studio
spawn_agent	Delegate a task to an autonomous sub-agent
create_tool	Create a reusable Python tool saved to the workspace
call_tool	Call a previously created tool by name
list_custom_tools	List all dynamically created tools
delete_tool	Delete a dynamic tool

---

Installation

Or during development:

---

Configuration

Open Settings → Plugins → high-perf-tools in LM Studio.

Setting	Default	Description
Workspace Path	(cwd)	All file operations are sandboxed here
Python Interpreter	(system python3)	Conda env name (e.g. misc) or full binary path
Allow Shell Commands	off	Enables run_command
Allow Python Execution	off	Enables run_python_code, run_python_file, run_tests
Allow C++ Compilation	off	Enables compile_cpp, run_binary, compile_and_run, cmake_build
Allow pip install

Python / Conda Environment

Set Python Interpreter to:

• Blank — uses system python3
• A conda env name (e.g. ml) — the plugin searches common conda prefix directories (~/miniconda3/envs, ~/anaconda3/envs, etc.) and falls back to conda run -n <name> if not found

All Python tools, pip, pytest, and dynamic tools share this single interpreter.

---

Agent Zero

Persistent Memory

The plugin stores memories as JSON at .agent_memory/memory.json inside the workspace. Memories survive across conversations. The model can tag and query them with memory_save / memory_recall.

Sub-agents (spawn_agent)

Sub-agents run an autonomous tool-call loop against LM Studio's OpenAI-compatible API.

RAM-aware model selection:

• One model loaded (low-RAM devices) → the same model is reused with a different system prompt. Zero extra RAM.
• Sub-Agent Model ID set and loaded → uses the dedicated sub-agent model.

The sub-agent has access to: read_file, write_file, list_directory, run_python_code, run_command, memory_save, memory_recall.

Set allow_tools=false for pure generation tasks (faster, no tool-call loop).

Dynamic Tools (create_tool / call_tool)

The model can write Python functions and register them as persistent tools:

Tools are saved as Python scripts in .agent_tools/ inside the workspace and can be called by name in future conversations. Arguments are passed as JSON via stdin; the function returns a string or JSON-serialisable value.

---

Error Handling

Every tool is wrapped in safe_impl(). When a tool throws, the model receives a structured JSON response it can parse and retry:

The prompt preprocessor injects tool-call rules on the first turn, reminding the model to output valid JSON and handle tool_error responses.

---

Security

• Workspace sandbox: all file paths are resolved and validated against the configured workspace root. Paths that escape (e.g. ../../etc/passwd) are rejected.
• Private IP blocking: fetch_url blocks requests to localhost, RFC 1918, and link-local addresses.
• Execution opt-in: shell, Python, C++, and pip tools are disabled by default and must be explicitly enabled in settings.

---

Requirements

• LM Studio with plugin support
• Node.js 18+
• lms CLI (npm install -g @lmstudio/lms)
• Optional: Python / conda, gcc/clang++, cmake, ruff, black, mypy, pytest