Notebooks · Specialist agents · 42

Power tools

Name: SHIPIT Agent
Author: SHIPIT

Three new tools from this release — each pairs naturally with one or more of the specialists from notebook 41.

Open in Colab View on GitHub 8 code cells · 0 with saved output

Its imports and arguments are checked against shipit-agent 3.0. The outputs were saved when it was written, for an earlier release.

In [ ]

python

from pathlib import Path
import sys

# Resolve the repo root whether this cell is running from ./notebooks
# or from the repo root — mirrors the existing 01–36 notebook series.
ROOT = (
    Path.cwd().resolve().parent
    if Path.cwd().name == "notebooks"
    else Path.cwd().resolve()
)
if str(ROOT) not in sys.path:
    sys.path.insert(0, str(ROOT))

from examples.run_multi_tool_agent import build_llm_from_env

# Bedrock Llama 4 Scout is the default model this library ships with
# (see `shipit_agent/config.py`). No model name required — the helper
# reads `AWS_REGION` / credentials from env and returns an adapter that
# works with every Autopilot / Agent class.
llm = build_llm_from_env("bedrock")
print("Bedrock LLM ready:", type(llm).__name__)

1. `computer_use` — desktop automation

computer_use is the library's Claude-Desktop-style primitive. It drives the LOCAL desktop via platform-native helpers:

macOS → screencapture + cliclick (or AppleScript fallback)
Linux → scrot / ImageMagick import + xdotool
Windows → PowerShell

In [ ]

python

from shipit_agent.tools.computer_use import ComputerUseTool

computer = ComputerUseTool()
print(
    "Schema enum:",
    computer.schema()["function"]["parameters"]["properties"]["action"]["enum"],
)
print("Output dir:", computer.output_dir)

1a. Take a screenshot

When run on your laptop, this actually captures the screen. In this notebook it'll succeed only if screencapture (macOS) / scrot (linux) / powershell (windows) is on PATH — otherwise you'll get a clear install hint.

In [ ]

python

from shipit_agent.tools.base import ToolContext

out = computer.run(ToolContext(prompt="demo"), action="screenshot")
print(out.text)

1b. Use it inside an Autopilot

For a tutorial / QA run, combine computer_use with a specialist and Autopilot: the agent drives the GUI, the runtime keeps it on track with a budget.

In [ ]

python

from shipit_agent import Agent
from shipit_agent.agents import AgentRegistry

registry = AgentRegistry()
def_ = registry.get("design-reviewer")
ui_agent = Agent(
    llm=llm,
    prompt=def_.prompt,
    tools=[computer],
    max_iterations=6,
    name="UI Reviewer",
)
# In your own environment, replace the prompt with something that
# drives your app (e.g. open Preferences, verify toggles, screenshot).
print("UI Reviewer agent ready:", ui_agent.name)

2. `research_brief` — one-call research primitive

Search the web, skim the top N pages, return a structured brief with numbered citations. No API key — uses DuckDuckGo HTML.

In [ ]

python

from shipit_agent.tools.research_brief import ResearchBriefTool

research = ResearchBriefTool()
out = research.run(
    ToolContext(prompt="demo"),
    query="uv vs poetry python package manager 2026",
    max_sources=4,
)
print(out.text)

Pass deep=True to also fetch each result page and embed a short summary. Use sparingly — one call fans out to N+1 HTTP requests.

In [ ]

python

out = research.run(
    ToolContext(prompt="demo"),
    query="SQLite vs DuckDB for analytics",
    max_sources=3,
    deep=True,
)
print(out.text[:1200])

3. `hubspot_ops` — CRM in one tool

One tool, many actions: search_contacts, search_companies, search_deals, get_*, create_contact, create_deal, add_note, list_owners. Auth is HUBSPOT_TOKEN env (private-app bearer token).

In [ ]

python

import os
from shipit_agent.tools.hubspot import HubspotTool

if os.environ.get("HUBSPOT_TOKEN"):
    hubspot = HubspotTool()
    out = hubspot.run(
        ToolContext(prompt="demo"), action="search_contacts", query="test", limit=5
    )
    print(out.text[:800])
else:
    print(
        "Set HUBSPOT_TOKEN to hit live HubSpot — schema still works for dry-run inspection:"
    )
    from shipit_agent.tools.hubspot import HubspotTool as H

    print(
        H(token=None).schema()["function"]["parameters"]["properties"]["action"]["enum"]
    )

4. Stacking tools in one long-running Autopilot

A single agent with all three tools plus a substantial budget can drive a surprisingly complex outbound-sales flow end to end: research accounts (research_brief), log activity (hubspot_ops), take screenshots for internal review (computer_use). Budgets make it safe to leave running.

In [ ]

python

from shipit_agent.autopilot import Autopilot, BudgetPolicy
from shipit_agent.deep import Goal

tools = [research, computer]  # add HubspotTool() when token is set
autopilot = Autopilot(
    llm=llm,
    goal=Goal(
        objective="Compare three vendors of time-series databases and produce a recommendation.",
        success_criteria=[
            "Each vendor has a 1-paragraph summary with citations",
            "A comparison table covering licensing + scale + pricing",
            "A 1-paragraph recommendation with the primary reason",
        ],
    ),
    tools=tools,
    budget=BudgetPolicy(
        max_iterations=8, max_seconds=600, max_tool_calls=30, max_dollars=3.0
    ),
)
print("Ready to run. autopilot.run(run_id='tsdb-shootout') when you're ready.")

Summary

Each tool is designed to be usable standalone, but the real leverage comes from handing them to an Autopilot with a thoughtful budget. That's the Claude-Desktop model: long-running, goal-directed, live-streamed, budget-gated.

Full series complete. Thanks for reading — now go build.

Power tools

Three new tools from this release — each pairs naturally with one or more of the specialists from notebook 41.

Its imports and arguments are checked against shipit-agent 3.0. The outputs were saved when it was written, for an earlier release.

In [ ]

python

from pathlib import Path
import sys

# Resolve the repo root whether this cell is running from ./notebooks
# or from the repo root — mirrors the existing 01–36 notebook series.
ROOT = (
    Path.cwd().resolve().parent
    if Path.cwd().name == "notebooks"
    else Path.cwd().resolve()
)
if str(ROOT) not in sys.path:
    sys.path.insert(0, str(ROOT))

from examples.run_multi_tool_agent import build_llm_from_env

# Bedrock Llama 4 Scout is the default model this library ships with
# (see `shipit_agent/config.py`). No model name required — the helper
# reads `AWS_REGION` / credentials from env and returns an adapter that
# works with every Autopilot / Agent class.
llm = build_llm_from_env("bedrock")
print("Bedrock LLM ready:", type(llm).__name__)

1. `computer_use` — desktop automation

computer_use is the library's Claude-Desktop-style primitive. It drives the LOCAL desktop via platform-native helpers:

macOS → screencapture + cliclick (or AppleScript fallback)
Linux → scrot / ImageMagick import + xdotool
Windows → PowerShell

In [ ]

python

from shipit_agent.tools.computer_use import ComputerUseTool

computer = ComputerUseTool()
print(
    "Schema enum:",
    computer.schema()["function"]["parameters"]["properties"]["action"]["enum"],
)
print("Output dir:", computer.output_dir)

1a. Take a screenshot

In [ ]

python

from shipit_agent.tools.base import ToolContext

out = computer.run(ToolContext(prompt="demo"), action="screenshot")
print(out.text)

1b. Use it inside an Autopilot

For a tutorial / QA run, combine computer_use with a specialist and Autopilot: the agent drives the GUI, the runtime keeps it on track with a budget.

In [ ]

python

from shipit_agent import Agent
from shipit_agent.agents import AgentRegistry

registry = AgentRegistry()
def_ = registry.get("design-reviewer")
ui_agent = Agent(
    llm=llm,
    prompt=def_.prompt,
    tools=[computer],
    max_iterations=6,
    name="UI Reviewer",
)
# In your own environment, replace the prompt with something that
# drives your app (e.g. open Preferences, verify toggles, screenshot).
print("UI Reviewer agent ready:", ui_agent.name)

2. `research_brief` — one-call research primitive

Search the web, skim the top N pages, return a structured brief with numbered citations. No API key — uses DuckDuckGo HTML.

In [ ]

python

from shipit_agent.tools.research_brief import ResearchBriefTool

research = ResearchBriefTool()
out = research.run(
    ToolContext(prompt="demo"),
    query="uv vs poetry python package manager 2026",
    max_sources=4,
)
print(out.text)

Pass deep=True to also fetch each result page and embed a short summary. Use sparingly — one call fans out to N+1 HTTP requests.

In [ ]

python

out = research.run(
    ToolContext(prompt="demo"),
    query="SQLite vs DuckDB for analytics",
    max_sources=3,
    deep=True,
)
print(out.text[:1200])

3. `hubspot_ops` — CRM in one tool

In [ ]

python

import os
from shipit_agent.tools.hubspot import HubspotTool

if os.environ.get("HUBSPOT_TOKEN"):
    hubspot = HubspotTool()
    out = hubspot.run(
        ToolContext(prompt="demo"), action="search_contacts", query="test", limit=5
    )
    print(out.text[:800])
else:
    print(
        "Set HUBSPOT_TOKEN to hit live HubSpot — schema still works for dry-run inspection:"
    )
    from shipit_agent.tools.hubspot import HubspotTool as H

    print(
        H(token=None).schema()["function"]["parameters"]["properties"]["action"]["enum"]
    )

4. Stacking tools in one long-running Autopilot

In [ ]

python

from shipit_agent.autopilot import Autopilot, BudgetPolicy
from shipit_agent.deep import Goal

tools = [research, computer]  # add HubspotTool() when token is set
autopilot = Autopilot(
    llm=llm,
    goal=Goal(
        objective="Compare three vendors of time-series databases and produce a recommendation.",
        success_criteria=[
            "Each vendor has a 1-paragraph summary with citations",
            "A comparison table covering licensing + scale + pricing",
            "A 1-paragraph recommendation with the primary reason",
        ],
    ),
    tools=tools,
    budget=BudgetPolicy(
        max_iterations=8, max_seconds=600, max_tool_calls=30, max_dollars=3.0
    ),
)
print("Ready to run. autopilot.run(run_id='tsdb-shootout') when you're ready.")

Summary

Full series complete. Thanks for reading — now go build.

Power tools

1. computer_use — desktop automation

1a. Take a screenshot

1b. Use it inside an Autopilot

2. research_brief — one-call research primitive

3. hubspot_ops — CRM in one tool

4. Stacking tools in one long-running Autopilot

Summary

Power tools

1. computer_use — desktop automation

1a. Take a screenshot

1b. Use it inside an Autopilot

2. research_brief — one-call research primitive

3. hubspot_ops — CRM in one tool

4. Stacking tools in one long-running Autopilot

Summary

1. `computer_use` — desktop automation

2. `research_brief` — one-call research primitive

3. `hubspot_ops` — CRM in one tool

1. `computer_use` — desktop automation

2. `research_brief` — one-call research primitive

3. `hubspot_ops` — CRM in one tool