GantryGraph Documentation

GantryGraph is an open-source Python library for building autonomous AI agents that operate the desktop, browser, filesystem, and external services via MCP. Built on LangGraph. MIT licensed.

Core idea

Every agent runs a four-step loop inside a GantryEngine:

observe → think → act → review
   ↑__________________________|

Observe — take a screenshot or accessibility snapshot of the current state.
Think — the LLM decides what tool to call next, or returns a final answer.
Act — tool calls execute; errors come back as recoverable messages.
Review — if tool calls were made, loop. Otherwise, done.

Quick example

from gantrygraph import GantryEngine
from gantrygraph.perception import WebPage
from gantrygraph.actions import BrowserTools
from langchain_anthropic import ChatAnthropic

web = WebPage(url="https://news.ycombinator.com", headless=True)

agent = GantryEngine(
    llm=ChatAnthropic(model="claude-sonnet-4-6"),
    perception=web,
    tools=[BrowserTools(web_page=web)],
    max_steps=20,
)

result = agent.run("Find the top 5 stories and return their titles and links.")
print(result)

Where to go next