Token-free AI tool calling - Access unlimited tools without context overhead
LTP allows AI agents to access unlimited MCP tools and Crafts without consuming tokens for tool definitions. Instead of loading tool schemas into context, AI uses simple CLI commands to discover and call tools on-demand.
Connect to any MCP server and call its tools directly:
Traditional MCP: 300 tools x 200 tokens = 60,000 tokens consumed
LTP: 0 tokens consumed - tools discovered on-demand
Crafts are reusable AI workflows - like Anthropic's Skills, but executable:
.ltp/crafts/
code-review/
CRAFT.md # Instructions for AI
execute.js # Automation script
execute.ts # Automation script
| Anthropic Skills | LTP Crafts |
|---|---|
| Static instructions | Instructions + Executable scripts |
| AI reads only | AI reads and executes |
| Single IDE | Works in any AI IDE |
LTP's Lazy Loading Pattern reduces token consumption by 75-90% compared to direct MCP calls.
| Method | Tokens (12 calls) | Cost | Reduction |
|---|---|---|---|
| Direct MCP | ~45,000 | ~$0.25 | - |
| LTP | ~7,500 | ~$0.07 | 83% |
Efficiency increases with more calls (converges at ~85-90%). See docs/effect.md for detailed analysis.
npm install -g ltp@latest# Add an MCP server
ltp add filesystem npx -y @modelcontextprotocol/server-filesystem
# List available tools
ltp list
# Search for tools
ltp search "read file"
# Call a tool
ltp call filesystem read_file --path /src/main.ts
# Get tool details
ltp info filesystem read_file| Command | Description |
|---|---|
ltp add <name> <cmd> |
Add an MCP server |
ltp remove <name> |
Remove an MCP server |
ltp servers |
List configured servers |
ltp list |
List servers with tool counts |
ltp list -v |
List tools with descriptions |
ltp list --schema |
List tools with parameter types |
ltp tools <server> |
List tool names for a server |
ltp search <query> |
Search for tools |
ltp info <server> <tool> |
Get tool details |
ltp call <server> <tool> |
Call a tool |
ltp call --file <path> |
Call with JSON file arguments |
ltp shell |
Interactive shell (persistent sessions) |
ltp crafts |
List available crafts |
ltp crafts-open |
Open crafts folder in explorer |
ltp craft <name> [--opts] |
Execute a craft with options |
ltp craft-install <url> |
Install/update craft from GitHub |
ltp craft-remove <name> |
Remove an installed craft |
ltp config open |
Open config file in editor |
ltp config path |
Show config file path |
ltp start |
Start daemon (fast mode) |
ltp stop |
Stop daemon |
ltp restart |
Restart daemon |
ltp status |
Check daemon status |
ltp sponsor |
Support the project |
Copy the contents of rule/AI_GLOBAL_RULES.md to your AI IDE's system prompt:
| IDE | Location |
|---|---|
| Antigravity | MEMORY[user_global] |
| Claude Code | ~/.claude/CLAUDE.md |
| Cursor | .cursorrules file |
| Other | System prompt settings |
This enables your AI agent to automatically discover and use LTP tools.
Config file location: ~/.ltp/config.json (Edit with ltp config open)
LTP supports two transport types for connecting to MCP servers:
| Transport | Description | Use Case |
|---|---|---|
stdio |
Spawns local process (default) | Local MCP servers, npx packages |
sse |
Connects via HTTP/SSE | Remote MCP servers, cloud APIs |
# stdio transport (default)
ltp add filesystem npx -y @modelcontextprotocol/server-filesystem
# SSE transport
ltp add remote-api --transport sse --url http://localhost:3000/sse{
"port": 7878,
"craftsDir": ".ltp/crafts",
"servers": {
"filesystem": {
"transport": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/dir"],
"env": {}
},
"github": {
"transport": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {
"GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
}
},
"remote-api": {
"transport": "sse",
"url": "https://api.example.com/mcp/sse"
}
}
}Crafts are reusable workflows with instructions and scripts.
Crafts are installed globally to ~/.ltp/crafts/ and shared across all projects.
# Install from GitHub repository root
ltp craft-install JuN-B-official/ltp
# Install from a subfolder (using --path option)
ltp craft-install JuN-B-official/ltp --path crafts/code-review
# Short option
ltp craft-install JuN-B-official/ltp -p crafts/code-review
# Remove a craft
ltp craft-remove code-reviewBrowse and discover community crafts at ltp.jun-b.com/crafts
You can also register your own crafts through the web interface.
Open crafts folder with ltp crafts-open
~/.ltp/crafts/
code-review/
CRAFT.md # Instructions
execute.js # Script
execute.ts # Script
For AI-friendly output:
ltp --json list
ltp --json call filesystem read_file --path /test.txtFor optimal token efficiency, AI agents should follow this workflow:
# 1. List available servers
ltp list
# Output: appwrite (143 tools), shopify (257 tools), ...
# 2. Get tool names for the needed server
ltp tools appwrite
# Output: list_databases, create_database, get_database, ...
# 3. Get details for the specific tool
ltp info appwrite create_database
# Output: description, input schema, usage example
# 4. Call the tool
ltp call appwrite create_database --databaseId mydb --name "My Database"Alternatively, get all tools with parameter types in one call:
ltp list --schema
# Output:
# filesystem (5 tools)
# read_file(path: string)
# write_file(path: string, content: string)
# ...For complex JSON arguments that are hard to escape:
# Create a JSON file with arguments
echo '{"path": "/src", "recursive": true}' > args.json
# Call with --file
ltp call filesystem list_directory --file args.json# Require confirmation for sensitive operations
ltp call --confirm filesystem delete_file --path /important.txtLove LTP? Consider supporting the project!
The Lazy Tool Protocol was created by JunHyeong Song @Song-JunHyeong.
See CONTRIBUTING.md.
Copyright 2025 JuN-B (JunHyeong Song)
Licensed under the Apache License, Version 2.0