LATS Server

A sophisticated code investigation agent that uses Language Agent Tree Search (LATS) with Monte Carlo Tree Search to systematically explore codebases and provide intelligent insights.

Features

• 🌳 Monte Carlo Tree Search: Systematic parallel exploration of solution space
• 🧠 Reasoning Transparency: Full chain-of-thought with gpt-oss model
• 💾 Persistent Memory: Learn from past investigations using langmem
• 🔍 Smart Code Analysis: AST-based structure analysis and dependency extraction
• 🚀 MCP Integration: Easy integration with Claude and other LLMs
• 📊 Pattern Recognition: Learns successful investigation patterns over time

Quick Start

Prerequisites

1. Python 3.9+
2. Ollama with gpt-oss model:

# Install Ollama (if not installed)
curl -fsSL https://ollama.com/install.sh | sh

# Pull the gpt-oss model
ollama pull gpt-oss

# Start Ollama server
ollama serve

Installation

# Clone the repository
git clone <repository-url>
cd lats

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Running the Server

# Make the server executable
chmod +x lats_mcp_server.py

# Run the MCP server
python lats_mcp_server.py

Integration with Claude

Add to your Claude MCP configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "lats": {
      "command": "python",
      "args": ["/absolute/path/to/lats_mcp_server.py"],
      "transport": "stdio"
    }
  }
}

Usage Examples

Basic Investigation

# Via MCP in Claude
"Investigate where error handling is implemented in the authentication module"

# Response includes:
# - Solution path with scored steps
# - File references with line numbers
# - Explored branches
# - Confidence score
# - Actionable suggestions

Quick File Analysis

# Analyze a specific file
"Analyze the structure of auth/login.py"

# Returns:
# - File content preview
# - Code structure (classes, functions)
# - Dependencies and imports

Parallel Search

# Search for multiple patterns simultaneously
"Search for 'login', 'authenticate', and 'session' in the codebase"

# Returns matches for each pattern with context

Available MCP Tools

investigate

Full LATS investigation of a task

• Args: task (str), max_depth (int), max_iterations (int), use_memory (bool)
• Returns: Solution path, file references, confidence score

get_status

Get current investigation status

• Returns: Task, status, progress, current branch

search_memory

Search past investigations

• Args: query (str), limit (int)
• Returns: Similar investigations with solutions

get_insights

Retrieve relevant insights

• Args: context (str)
• Returns: List of relevant insights

analyze_file

Quick single-file analysis

• Args: file_path (str)
• Returns: Content, structure, dependencies

parallel_search

Search multiple patterns in parallel

• Args: patterns (List[str]), directory (str)
• Returns: Matches for each pattern

How LATS Works

1. Tree Search Process

Root Node
├── Action 1 (Score: 6.5)
│   ├── Action 1.1 (Score: 7.8) ← Best path
│   └── Action 1.2 (Score: 5.2)
└── Action 2 (Score: 4.3)
    └── Action 2.1 (Score: 3.9)

2. Node Selection

Uses Upper Confidence Bound (UCT) to balance:

• Exploitation: Choose high-scoring paths
• Exploration: Try less-visited branches

3. Reflection & Scoring

Each action is evaluated on:

• Relevance to task (0-10 scale)
• Information quality
• Progress toward solution

4. Memory & Learning

• Stores successful investigations
• Extracts action patterns
• Provides suggestions for similar tasks

Configuration

Edit LATSConfig in lats_core.py:

class LATSConfig:
    model_name = "gpt-oss"          # Ollama model
    base_url = "http://localhost:11434"  # Ollama URL
    temperature = 0.7                # Model temperature
    max_depth = 5                    # Max tree depth
    max_iterations = 10              # Max search iterations
    num_expand = 5                   # Actions per expansion
    c_param = 1.414                  # UCT exploration parameter
    min_score_threshold = 7.0        # Solution threshold

Architecture

┌─────────────────┐
│   MCP Client    │
│    (Claude)     │
└────────┬────────┘
         │ MCP Protocol
┌────────▼────────┐
│  FastMCP Server │
└────────┬────────┘
         │
┌────────▼────────┐
│  LATS Algorithm │
├─────────────────┤
│ • Tree Search   │
│ • Node Selection│
│ • Reflection    │
└────────┬────────┘
         │
┌────────▼────────────┐
│   Core Components   │
├────────┬────────────┤
│Filesystem│  Memory  │
│  Tools   │ Manager  │
└──────────┴──────────┘
         │
┌────────▼────────┐
│     Ollama      │
│   (gpt-oss)     │
└─────────────────┘

Development

Running Tests

# Run unit tests
python -m pytest tests/

# Run with coverage
python -m pytest --cov=. tests/

Adding New Tools

1. Add tool function to filesystem_tools.py
2. Register in create_filesystem_tools()
3. Update MCP server if needed

Extending Memory

1. Add namespace in MemoryManager.__init__
2. Create storage/retrieval methods
3. Integrate with investigation flow

Troubleshooting

Ollama Connection Issues

# Check Ollama is running
curl http://localhost:11434/api/tags

# Verify model is available
ollama list | grep gpt-oss

Memory Store Errors

• Check write permissions in directory
• Verify langmem is properly installed
• Review error namespace for details

Tool Execution Failures

• Check file permissions
• Verify path existence
• Review size limits (1MB max)

Performance Tips

1. Adjust max_depth: Lower for faster results
2. Limit iterations: Reduce for quicker investigations
3. Use memory: Leverages past investigations
4. Parallel search: Batch multiple queries
5. Target searches: Provide specific directories

Contributing

1. Fork the repository
2. Create feature branch
3. Add tests for new features
4. Update documentation
5. Submit pull request

License

MIT License - See LICENSE file for details

Acknowledgments

• LangChain/LangGraph for agent framework
• Anthropic for MCP protocol
• OpenAI for gpt-oss model
• langmem for memory management