Netweb SDK

Netweb SDK is an enterprise-grade agent engineering framework for building, orchestrating, and deploying AI assistants with custom tools and multiple LLM providers.

Key capabilities:

• Custom tool integration with the @nw_tool decorator
• Composable agents via create_agent
• Deterministic, structured tool invocation (JSON/structured calls)
• Multiple built-in provider adapters (OpenAI, Anthropic, Google, open-source)
• Pluggable memory stores (ConversationMemory, VectorMemory, PersistentMemory)
• Tracing integration (LangSmith-compatible) via enable_tracing

Why use Netweb SDK

Netweb SDK is intended for production AI applications that require:

• Reliable tool invocation from generated outputs
• Reusable agent components and configurable system prompts
• Provider-agnostic integrations for deploying across clouds and local runtimes

Installation

Install from PyPI:

pip install netweb-sdk

Or install from source:

pip install .

Quick Start

The SDK exports common helpers from sdk (see sdk/__init__.py): create_agent, get_tool, get_tools, list_tools, nw_tool, provider classes, memory helpers, and tracing utilities.

1. Define a custom tool with @nw_tool:

from sdk import nw_tool

@nw_tool("add", description="Add two integers together")
def add(a: int, b: int) -> int:
    return a + b

2. Create an agent. Note: current create_agent signature accepts name, role, llm, tools, memory (bool or instance), input_schema, output_schema, prompt_path, system_prompt, retries, and retry_delay.

from sdk import create_agent, OpenAIProvider

llm = OpenAIProvider(model="gpt-4o-mini")

agent = create_agent(
    name="calculator_agent",
    role="calculator",
    llm=llm,
    tools=[add],
    system_prompt="You are a precise calculator. Use the add tool for addition operations.",
    memory=True,
)

3. Run the agent (sync or streamed):

resp = await agent.nw_run("Add 5 and 3 together")
print(resp.output)

# Streamed
async for chunk in agent.nw_run_streamed("Tell me a short joke"):
    print(chunk, end="", flush=True)

Providers

The SDK includes adapters for multiple LLM providers. Examples exported from sdk:

• OpenAIProvider
• AnthropicProvider
• GoogleProvider
• OpenSourceProvider (HuggingFace, Ollama, local runtimes)

Provider classes wrap the API surface for sync/async generation and streaming and are pluggable when creating agents.

Memory

Memory implementations available in the SDK:

• ConversationMemory: in-memory conversation history
• VectorMemory: vector-store backed memory for retrieval-augmented generation
• PersistentMemory: durable stores (e.g., redis/postgres checkpointers supported in runtime)

Pass memory=True to create_agent to enable default conversation memory (a ConversationMemory instance).

Tools API

Tool utilities exported from sdk:

• nw_tool: decorator for registering tools
• get_tools(names: list[str]): fetch tool instances by name
• get_tool(name: str): fetch a single tool
• list_tools(): list available tool metadata

Tools are executed by the agent via ToolRunner and are registered in the ToolRegistry on agent creation.

Tracing

Enable LangSmith-compatible tracing using:

from sdk.tracing import enable_tracing
enable_tracing(project_name="my_project")

Tracing hooks into execution to capture runs, tool calls, and metadata.

Prompts and Templates

Prompts are managed via abstractions.agents.prompts with utilities for loading prompt files, templating, and prompt composition. You can pass prompt_path to create_agent to use a custom prompt file, or provide system_prompt directly.