Hamilton Mcp

Or use programmatically:

from hamilton.plugins.h_mcp import get_mcp_server

mcp = get_mcp_server()
mcp.run()

Workflow: The Golden Path

Always follow this sequence when building Hamilton DAGs interactively:

ask user -> capabilities -> scaffold -> validate -> visualize -> correct -> execute

Step 1: Ask the User Which Libraries They Use

Before calling any tool, ask the user which data libraries they use (pandas, numpy, polars, etc.). Then pass their answer as preferred_libraries to hamilton_capabilities and hamilton_scaffold. This ensures scaffolds match the user's project, not the server's environment.

// Example: user says "I use pandas"
// Tool call: hamilton_capabilities(preferred_libraries=["pandas"])
{
  "libraries": {
    "pandas": true,
    "numpy": true,
    "polars": false,
    "graphviz": true
  },
  "available_scaffolds": [
    "basic", "basic_pure_python", "config_based",
    "data_pipeline", "parameterized"
  ]
}

Decision rules:

• If user says pandas: use pandas-based scaffolds and DataFrame/Series types
• If user has no preference or only uses built-in types: use basic_pure_python scaffold and int/float/str/dict types
• If graphviz is available: use hamilton_visualize to show the DAG structure
• Never generate code that imports libraries the user hasn't stated they use

Step 2: Scaffold a Starting Point

Use hamilton_scaffold with a pattern name from the capabilities response:

Step 3: Validate Before Executing

Always validate before executing. hamilton_validate_dag compiles the DAG without running it, catching:

• Syntax errors
• Missing dependencies (parameter names that don't match any function)
• Type annotation issues
• Circular references

// Success response
{
  "valid": true,
  "node_count": 5,
  "nodes": ["cleaned", "feature_a", "feature_b", "raw_data", "result"],
  "inputs": ["data_path"],
  "errors": []
}

// Failure response
{
  "valid": false,
  "node_count": 0,
  "nodes": [],
  "inputs": [],
  "errors": [{"type": "SyntaxError", "message": "...", "detail": "line 5"}]
}

Self-correction loop: If validation fails, read the error, fix the code, and validate again. Do not proceed to execution until validation passes.

Step 4: Visualize the DAG (if graphviz available)

hamilton_visualize returns DOT graph source. Use this to:

• Confirm dependency structure matches intent
• Identify unexpected connections
• Explain the pipeline to the user

Step 5: Explore Node Details

hamilton_list_nodes returns structured info for every node:

• Name, output type, documentation
• Whether it's an external input (must be provided at runtime)
• Required and optional dependencies

Use this to understand what inputs the DAG needs before execution.

Step 6: Execute

hamilton_execute runs the DAG with provided inputs and returns results. Key parameters:

• code: The full Python source
• final_vars: List of node names to compute (only these and their dependencies run)
• inputs: Dict of external input values
• timeout_seconds: Safety limit (default 30s)

WARNING: This executes arbitrary Python code. Always validate first.

Error Handling & Self-Correction

Common Errors and Fixes

"No module named 'X'" The code imports a library that isn't installed. Call hamilton_capabilities to check availability, then rewrite without the missing library.

"Missing dependencies: ['node_name']" A function parameter doesn't match any function name or external input. Either:

1. Add a function with that name, or
2. Include it in inputs when executing

"Execution timed out after Ns" The code takes too long. Reduce data size, simplify computation, or increase timeout_seconds.

Validation passes but execution fails Validation checks structure, not runtime behavior. Common causes:

• Missing input values at execution time
• Runtime exceptions in function bodies (division by zero, key errors)
• Library-specific errors (e.g., column not found in DataFrame)

Retry Strategy

1. If a tool returns an error, fix the issue in code and retry once
2. If the same error recurs, explain the issue to the user and ask for guidance
3. Never retry more than twice on the same error

Tool Reference

Environment Fallbacks

If the MCP server is not running: Fall back to CLI:

# Validate a module
python -c "from hamilton import driver; import my_module; dr = driver.Builder().with_modules(my_module).build(); print('Valid!')"

If Hamilton is not installed: Provide the user with installation instructions:

uvx --from "apache-hamilton[mcp]" hamilton-mcp   # Run via uvx (add --with <lib> as needed)
pip install "apache-hamilton[mcp]"              # Or install directly

Success Criteria

A successful MCP interaction produces:

1. Code that passes hamilton_validate_dag with zero errors
2. All external inputs identified via hamilton_list_nodes
3. Execution results returned from hamilton_execute
4. The user understands the DAG structure (via visualization or node listing)

Additional Resources

• For core Hamilton patterns: use /hamilton-core
• For scaling with async/Spark: use /hamilton-scale
• For LLM workflow patterns: use /hamilton-llm
• For observability: use /hamilton-observability
• Hamilton documentation: hamilton_get_docs("overview")