langterm

Langterm

Langterm translates natural language to executable shell commands using local AI models through Ollama. It works on Windows, macOS, and Linux.

🚀 New in v1.2.0: Project Memory System for location-aware context enhancement, User-Defined Hooks system for reusable command templates, plus MCP (Model Context Protocol) support with intelligent routing that can execute MCP tools directly, generate terminal commands, or use hybrid approaches for optimal results.

✨ v1.0.1: Enhanced security with dangerous command detection and confirmation system to keep your system safe.

Prerequisites

• Node.js 18+ - Required to run langterm
• Ollama - Install from https://ollama.com
• An Ollama model - Pull one before first use:

  ollama pull codestral:22b
  # or for a lighter/faster model:
  ollama pull deepseek-coder:6.7b
  

Docker/Remote Ollama Setup

If Ollama is running in Docker or on a remote server:

# Set the Ollama URL environment variable
export OLLAMA_URL=http://host.docker.internal:11434  # For Docker
export OLLAMA_URL=http://192.168.1.100:11434         # For remote server

Installation

Using npx (no installation needed)

npx langterm

Global installation

npm install -g langterm

Usage

First Run

When you run langterm for the first time, it will:

• Check if Ollama is running
• Show available models
• Let you select your preferred model
• Save your choice for future use

Basic Usage

# Interactive mode
langterm

# Direct command
langterm list all files larger than 100MB

# With quotes for complex commands
langterm "find all Python files modified in the last week"

Examples

langterm show disk usage sorted by size
langterm find process running on port 8080
langterm create tar archive excluding node_modules
langterm "convert all PNG images to JPG in current directory"
langterm extract audio from video.mp4 as mp3

Options

• langterm --setup - Reconfigure model selection
• langterm --model <name> - Use a specific model for this run
• langterm --mcp-setup - Configure MCP servers for enhanced context
• langterm --mcp-status - Show MCP connection status
• langterm --mcp-enable/--mcp-disable - Toggle MCP integration
• langterm --hooks-create <name> - Create a new hook template
• langterm --hooks-list - List all available hooks
• langterm --hooks-edit <name> - Edit an existing hook
• langterm --hooks-delete <name> - Delete a hook
• langterm --hooks-search <term> - Search hooks by name or content
• langterm /hookname - Use a hook template
• langterm --remember <info> - Save project information
• langterm --recall - Show saved project memory
• langterm --forget - Clear project memory
• langterm --memory-status - Show memory status
• langterm --help - Show help message
• langterm --version - Show version

How It Works

• You describe what you want in plain English
• Project Memory (automatic): Langterm loads project-specific context and learning patterns
• Enhanced Context (if MCP enabled): Langterm gathers relevant context from connected servers
• Langterm sends your request + context to your local Ollama model
• The AI generates the appropriate shell command with enhanced accuracy
• Security check: Langterm analyzes the command for dangerous patterns
• You review the command and confirm execution based on the security level
• Learning (automatic): Successful commands are recorded for future reference

MCP (Model Context Protocol) Support

What is MCP?

MCP allows Langterm to connect to external servers for enhanced context awareness. NEW in v1.2.0: MCP-first intelligent routing that prioritizes MCP tools when available:

• MCP-First Approach: Always attempts to match user requests with available MCP tools first
• Intelligent Matching: Analyzes tool capabilities against user input without hardcoded patterns
• Terminal Fallback: Only generates terminal commands when no suitable MCP tools are found
• Smart Tool Selection: Automatically selects the most relevant MCP tools based on request

Quick Start with MCP

• Setup MCP servers:

  langterm --mcp-setup
  

• Check connection status:

  langterm --mcp-status
  

• Experience intelligent routing:

  # File operations → Direct MCP tool execution
  langterm "read the package.json file"
  # Result: Uses MCP filesystem tool directly (with confirmation)
  
  # System operations → Terminal command generation  
  langterm "find files larger than 100MB"
  # Result: Generates optimized find command
  
  # Complex tasks → Hybrid approach
  langterm "analyze Python imports in project files"
  # Result: MCP finds files + generates analysis commands
  

Example MCP Benefits

Before MCP:

$ langterm "show me the main configuration file"
# Result: ls *.conf *.config

With MCP (filesystem server providing project context):

$ langterm "show me the main configuration file"
# Enhanced context: "Project uses .env, package.json, docker-compose.yml"
# Result: cat package.json

Supported MCP Server Types

• Local servers (stdio): Filesystem access, Git integration, project tools
• Remote servers (HTTP/SSE): Web APIs, cloud services, databases

Security with MCP

• ✅ Same security validation: All MCP-enhanced commands go through identical security checks
• ✅ No privilege escalation: MCP provides context, not execution permissions
• ✅ User control: You configure which servers to connect to
• ✅ Graceful fallback: If MCP fails, commands work without enhanced context

Location-Based Memory System

NEW in v1.2.0: Langterm saves context for any directory, enhancing commands with location-specific information.

How Memory Works

Langterm:

• Saves memory per directory - Each location can have its own .langterm-memory.json
• Works anywhere - Not limited to "projects", works in any directory
• Learns from successful commands - Improves future suggestions
• Auto-detects context - Recognizes project types, languages, and build systems

Quick Start with Memory

• Save information for current directory:

  langterm --remember "This is a React TypeScript project using Vite and Tailwind"
  

• Use enhanced commands:

  # Memory automatically provides context about your React/TypeScript setup
  langterm "run tests"
  # Result: npm test (or yarn test, based on your project)
  
  langterm "build for production" 
  # Result: npm run build (optimized for your Vite setup)
  

• View your project memory:

  langterm --recall
  

Automatic Learning

Langterm learns from your usage patterns:

• Successful commands are recorded for future reference
• Similar requests get suggestions from your history
• Project preferences are detected and remembered

Example Memory Benefits

Before Memory:

$ langterm "run tests"
# Generic result: npm test

With Memory (React TypeScript project):

$ langterm "run tests"
# Enhanced context: "React TypeScript project using Vite, prefers npm"
# Smart result: npm run test:coverage

Memory Commands

# Save information about your project
langterm --remember "Uses microservices architecture with Docker"

# View all saved memory for current project
langterm --recall

# Clear memory for current project
langterm --forget

# Check memory status across projects
langterm --memory-status

What Gets Detected Automatically

• Project Type: Node.js, Python, Rust, Go, Java, PHP, Ruby, Docker, etc.
• Primary Language: Based on file extensions in your project
• Build System: npm, yarn, cargo, make, gradle, maven, etc.
• Framework Patterns: React, Vue, Express, FastAPI, etc. (from file structure)

Memory Safety

• ✅ Local storage only: Memory files stay in your project directories
• ✅ Size limits: Maximum 50KB per project to prevent bloat
• ✅ Security scanning: All memory content is validated for safety
• ✅ Easy cleanup: Delete .langterm-memory.json to reset

User-Defined Hooks

NEW in v1.2.0: Create reusable command templates stored as markdown files that you can invoke with simple /hookname syntax.

How Hooks Work

Instead of typing the same long descriptions repeatedly, create hooks that store your frequently-used commands:

• Create a hook with a descriptive name
• Store natural language in a .md file
• Invoke with /hookname syntax
• Langterm processes the hook content as if you typed it

Quick Start with Hooks

• Create your first hook:

  langterm --hooks-create backup
  # Enter: "create a compressed backup of the src directory with today's date"
  

• Use your hook:

  langterm /backup
  # Same as: langterm "create a compressed backup of the src directory with today's date"
  

• List all hooks:

  langterm --hooks-list
  

Hook Management

# Create a new hook template
langterm --hooks-create deploy
langterm --hooks-create git-status
langterm --hooks-create cleanup

# Edit existing hooks
langterm --hooks-edit backup

# Delete hooks you no longer need  
langterm --hooks-delete old-hook

# Search through your hooks
langterm --hooks-search git

Hook Examples

backup.md:

create a compressed backup of the src/ directory with today's date in the filename

deploy.md:

build the project, run tests, and deploy to production server using rsync

git-status.md:

show git status with information about uncommitted changes and current branch

cleanup.md:

remove all .tmp files, clear npm cache, and delete node_modules in nested directories

monitor.md:

show system resource usage including CPU, memory, and disk space with human-readable format

Usage Examples

# Instead of typing this every time:
langterm "create a compressed backup of the src/ directory with today's date in the filename"

# Create a hook and use it:
langterm --hooks-create backup
langterm /backup

# More examples:
langterm /deploy      # Deploy your application
langterm /git-status  # Check git status with details
langterm /cleanup     # Clean up temporary files
langterm /monitor     # Check system resources

Hook Storage

Hooks are stored as markdown files in ~/.langterm/hooks/:

~/.langterm/hooks/
├── backup.md
├── deploy.md  
├── git-status.md
└── cleanup.md

Hook Safety

• ✅ Name validation: Only alphanumeric, dashes, and underscores allowed
• ✅ Content validation: Basic checks for dangerous patterns in templates
• ✅ File size limits: Maximum 10,000 characters per hook
• ✅ Security scanning: Hook content is scanned for potentially dangerous patterns

Security Features

Langterm includes comprehensive safety measures to protect against dangerous commands:

🔴 Dangerous Commands

Commands that could seriously damage your system require typing "YES I AM SURE" exactly:

• File system destruction (rm -rf /, format C:)
• Fork bombs and infinite loops
• System file corruption
• Permission changes that compromise security
• Network security bypasses
• System shutdown commands

🟡 Warning Commands

Potentially risky commands require typing "yes" to confirm:

• Commands requiring elevated privileges (sudo)
• Recursive force deletions (rm -rf)
• File overwrites (> redirections)
• Executing remote scripts (curl | sh)

🟢 Safe Commands

Normal commands only require pressing Enter to execute:

• File listings, searches, and reads
• Process management
• Archive operations
• Safe file operations

Example security prompts:

$ langterm "delete everything on my computer"
⚠️ DANGEROUS COMMAND DETECTED ⚠️
Risk: Recursive deletion of root directory
Command: rm -rf /
This command could cause serious damage to your system!
Type "YES I AM SURE" to continue, or anything else to cancel:

Configuration

Langterm saves your model preference in ~/.langtermrc. You can:

• Run langterm --setup to change models
• Edit the file directly
• Use --model flag to temporarily use a different model

Troubleshooting

"Ollama is not running"

Langterm will now provide helpful hints when Ollama isn't running:

❌ Ollama is not running!
💡 Hint: Is Ollama running? Try: ollama serve

Start Ollama with:

ollama serve

"No models found"

Pull a model first:

ollama pull codestral:22b

Command generation issues

• Try being more specific in your description
• Some models work better than others for command generation
• Codestral and DeepSeek Coder models are recommended

Security warnings

• If you see dangerous command warnings, double-check what you're asking for
• The AI might misinterpret your request - review the generated command carefully
• You can always cancel with Ctrl+C if something doesn't look right

Privacy

All processing happens locally on your machine. No data is sent to external servers.

License

MIT