Catalog Overview

shipit-agent ships with 25+ production-ready built-in tools organized into four categories. Every tool is opt-in (use only what you need), discoverable via tool_search, and follows the same Tool protocol so you can mix built-ins with custom tools without ceremony.

Quick navigation

Full alphabetical index

How tools are loaded

Three ways, pick whichever fits:

Way 1 — Everything at once:

from shipit_agent import Agent
from shipit_agent.llms import OpenAIChatLLM

agent = Agent.with_builtins(llm=OpenAIChatLLM(model="gpt-4o-mini"))
# All 25+ built-in tools registered automatically

Way 2 — Selective:

from shipit_agent import Agent, WebSearchTool, OpenURLTool, ToolSearchTool
from shipit_agent.llms import OpenAIChatLLM

agent = Agent(
    llm=OpenAIChatLLM(model="gpt-4o-mini"),
    tools=[WebSearchTool(),
        OpenURLTool(),
        ToolSearchTool(),],
)

Way 3 — Mix built-ins with custom tools:

from shipit_agent import Agent, FunctionTool, get_builtin_tools
from shipit_agent.llms import OpenAIChatLLM

def my_pricing_lookup(sku: str) -> str:
    """Look up internal product pricing by SKU."""
    return get_pricing_from_db(sku)

agent = Agent(
    llm=OpenAIChatLLM(model="gpt-4o-mini"),
    tools=[*get_builtin_tools(llm=OpenAIChatLLM(model="gpt-4o-mini")),
        FunctionTool.from_callable(my_pricing_lookup),],
)

How tools surface in events

Every tool call emits three events the runtime guarantees in this order:

tool_called      → "Tool called: web_search"           (before execution)
tool_completed   → "Tool completed: web_search"        (on success)
tool_failed      → "Tool failed: web_search"           (on error or unknown tool)

Both tool_called and tool_completed carry the full arguments and output in their event payloads. See the Event types reference for the complete schema.

Bedrock tool-pairing safety

The runtime guarantees that every toolUse block in an assistant turn gets a paired toolResult block in the next user turn — even when:

• A tool raises a non-retryable exception → synthetic error result is appended
• The model hallucinates an unregistered tool name → synthetic "Error: tool X is not registered" result
• The planner runs before the first LLM call → output is injected as a user-role context message, not an orphan tool-role result

This invariant is what makes multi-iteration tool loops on AWS Bedrock work reliably. See architecture for the full story.

Adding your own tool

Every tool implements three things: name, schema(), and run(context, **kwargs). See the custom tools guide for the full template, or the worked example in examples/05_custom_tool.py.

Where these are defined

Every built-in tool lives at shipit_agent/tools/<tool_name>/<tool_name>_tool.py. The implementation is small, readable, and a good template for your own tools — the Web search tool is ~150 lines including all five provider backends.

Related

• Custom tools guide — build a new tool from scratch
• Tool search — semantic discovery across many tools
• Architecture — how tools fit into the runtime loop
• Event types — tool_called / tool_completed / tool_failed payloads