Skip to content

README.md

Latest commit

 

History

History
139 lines (105 loc) · 4.11 KB

README.md

File metadata and controls

139 lines (105 loc) · 4.11 KB
Raw

License mloda Python

PTC + mloda Demo

3 ways to connect an LLM to your data. Same question, same two tools, very different plumbing.

The 3 Approaches

Loop (traditional agentic loop): Your code dispatches every tool call in a while-loop. Full control, high token cost.

Bash (Claude Code style): One claude -p call with Bash access. Claude writes Python and runs it directly. No tool definitions needed.

PTC (Programmatic Tool Calling): Claude writes Python in a sandboxed container, calls your tools as async functions. Results stay in the sandbox, only print() output goes back.

The Code

1/3: The Agentic Loop

Every tool call is a round-trip. Every result lands in context.

while True:
    response = client.messages.create(
        model="claude-sonnet-4-6",
        tools=TOOLS,
        messages=messages,
    )
    if response.stop_reason == "end_turn":
        break

    for block in response.content:
        if block.type == "tool_use":
            result = handle_tool(block.name, block.input)
            #        ^^^^ full result goes into context window

2/3: The Bash Shortcut

No tool loop. Claude just writes code. Fast, but broad permissions.

subprocess.run(
    ["claude", "-p", "--allowedTools", "Bash"],
    input="Use mloda to fetch employee data and find top earners",
)
# Claude writes + runs Python. No tool definitions needed.
# But: full shell access, limited observability.

3/3: Programmatic Tool Calling

Sandboxed. Tool results stay in the sandbox. Only print() goes back.

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16384,
    tools=[
        {"type": "code_execution_20260120", "name": "code_execution"},
        {
            "name": "run_features",
            "description": "Fetch data for given feature names. Returns CSV.",
            "input_schema": {
                "type": "object",
                "properties": {
                    "feature_names": {"type": "array", "items": {"type": "string"}}
                },
                "required": ["feature_names"],
            },
            "allowed_callers": ["code_execution_20260120"],  # callable from sandbox
        },
    ],
    messages=messages,
)

# Claude writes and runs this inside the sandbox:
#
#   csv = await run_features(feature_names=["employee_id", "salary", "department"])
#   df = pd.read_csv(io.StringIO(csv))
#   top3 = df.nlargest(3, "salary")
#   print(top3.to_string())   # only print() output goes back to Claude

Quick Start

uv venv && source .venv/bin/activate
uv sync --all-extras

Run all 3 approaches:

python demo.py

Run a single approach:

python demo.py loop
python demo.py bash
python demo.py ptc

loop and ptc require an ANTHROPIC_API_KEY. bash requires claude CLI installed.

How It Works

All 3 approaches use the same two tools:

  • discover_features -- list available mloda feature groups and their features
  • run_features -- fetch data for given feature names via mloda.run_all()

The data comes from a hardcoded employee dataset (id, department, salary, experience, performance score). Each approach answers the same 3 questions about this data.

The difference is only in how the model reaches the tools.

Project Structure

demo.py                           # all 3 approaches in one file
ptc_mloda_demo/
  feature_groups/sample_data/     # employee dataset (FeatureGroup)
  extenders/observability/        # observability extender
tests/
  test_mloda_imports.py

Checks

tox

Links