Compact Mode Guide¶
Compact mode exposes only 4-5 core tools instead of 1000+ tools, reducing context window usage by ~99% while maintaining full functionality.
What is Compact Mode?¶
Compact mode is a context window optimization that exposes only 4-5 core discovery tools instead of listing all 1000+ tools. All tools remain accessible through the execute_tool function.
Why Use Compact Mode?¶
The Problem: AI assistants have limited context windows. Listing 1000+ tools consumes significant context space, leaving less room for:
Your research questions
Tool outputs and data
Conversation history
Analysis and reasoning
The Solution: Compact mode exposes only 4-5 discovery tools. The AI uses these to find and execute the specific tools it needs.
Impact:
99% reduction in exposed tools (4-5 vs 1000+)
Full functionality - All 1000+ tools still accessible via
execute_toolBetter AI reasoning - More context available for analysis
Faster responses - Less token processing overhead
When to use:
✅ Always recommended for MCP integration with Claude, Cursor, ChatGPT ✅ AI agent workflows with long conversations ✅ Complex multi-step research tasks ✅ Working with large datasets or outputs
When to skip:
❌ Python API usage (context windows don’t apply) ❌ Single-use scripts or batch processing ❌ When you need to see all available tools at once
Core Tools¶
Compact mode exposes 4 core discovery tools, plus optionally find_tools if search is enabled (default):
``list_tools`` - List available tools (names, categories, etc.)
``grep_tools`` - Search tools by text/regex pattern
``get_tool_info`` - Get tool information (description or full definition)
``execute_tool`` - Execute any ToolUniverse tool by name
``find_tools`` - AI-powered tool discovery using natural language queries (optional, enabled by default via
search_enabled=True)
Quick Start¶
Command Line¶
# STDIO mode (for Claude Desktop)
tooluniverse-smcp-stdio --compact-mode
# HTTP mode
tooluniverse-smcp-server --compact-mode --port 8000
Claude Desktop Configuration¶
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"tooluniverse-compact": {
"command": "python",
"args": [
"-m", "tooluniverse.smcp_server",
"--transport", "stdio",
"--compact-mode"
],
"env": {
"FASTMCP_NO_BANNER": "1",
"PYTHONWARNINGS": "ignore"
}
}
}
}
Gemini CLI Configuration¶
Add to ~/.gemini/settings.json or project .gemini/settings.json:
{
"mcpServers": {
"tooluniverse": {
"command": "uv",
"args": [
"--directory",
"/path/to/tooluniverse-env",
"run",
"tooluniverse-smcp-stdio",
"--compact-mode"
]
}
}
}
Why Compact Mode for Gemini CLI?
500 tool limit: Gemini CLI has a 500 tool limit per MCP server
Minimal context usage: Compact mode exposes only 4-5 tools, well within limits
Full functionality: All 1000+ tools still accessible via
execute_toolProgressive disclosure: Discover tools on demand using
list_tools,grep_tools, andget_tool_info
Note
For Gemini CLI, you can also use the gemini-essential.yaml Space configuration which provides ~400-450 essential tools, staying within the 500 tool limit while providing direct access to commonly used tools. See building_ai_scientists/gemini_cli for details.
Usage¶
In Claude Desktop, just configure the server and start using tools. Claude will automatically discover and call them.
Typical workflow:
1. Use list_tools(mode="names") to see available tools
2. Use grep_tools(pattern="...") to search for tools
3. Use get_tool_info(tool_names="...") to get details
4. Use execute_tool(tool_name="...", arguments={...}) to execute tools
Comparison¶
Feature |
Normal Mode |
Compact Mode |
|---|---|---|
Tools Exposed |
~1000 tools |
4-5 tools (4 core + find_tools if search enabled) |
Context Usage |
High |
Low (99% reduction) |
Functionality |
Full |
Full (via execute_tool) |
When to Use¶
Use Compact Mode when: - Working with AI agents (Claude Desktop, Gemini CLI, etc.) - Context window is limited - Working with Gemini CLI (500 tool limit) - You want minimal context usage
Use Normal Mode when: - Context window is not a concern - You want direct access to all tools - Working with platforms without tool limits
Alternative for Gemini CLI:
- Use gemini-essential.yaml Space configuration for ~400-450 essential tools
- Provides direct tool access while staying within 500 tool limit
- See building_ai_scientists/gemini_cli for configuration details
Examples¶
See examples/compact_mode/ for complete examples.