Griptape — Modular Python AI Agent Framework

Griptape Structures & Patterns

Three Core Structures

from griptape.structures import Agent, Pipeline, Workflow

# 1. Agent — Single task with tools
agent = Agent(tools=[WebScraper(), Calculator()])
agent.run("What's 15% tip on a $85 dinner?")

# 2. Pipeline — Sequential multi-step tasks
pipeline = Pipeline()
pipeline.add_tasks(research_task, analyze_task, report_task)
pipeline.run("Analyze Q4 sales data")

# 3. Workflow — Parallel + sequential DAG
workflow = Workflow()
workflow.add_tasks(task_a, task_b, task_c)  # a,b run in parallel
task_c.add_parents([task_a, task_b])         # c waits for both
workflow.run()

Off-Prompt Processing

Griptape's signature feature — keep sensitive data out of the LLM:

from griptape.structures import Agent
from griptape.drivers import LocalVectorStoreDriver
from griptape.engines import VectorQueryEngine

# Data stays in local vector store, never sent to LLM
vector_store = LocalVectorStoreDriver()
query_engine = VectorQueryEngine(vector_store_driver=vector_store)

agent = Agent(
    tools=[
        # Tool processes data locally, sends only summaries to LLM
        VectorStoreTool(query_engine=query_engine, off_prompt=True)
    ]
)
agent.run("What are the key risks in our financial report?")
# LLM sees: "Based on the retrieved context: [summary]"
# LLM never sees: raw financial data

Tool System

Pre-built tools for common agent tasks:

Build custom tools:

from griptape.tools import BaseTool
from griptape.artifacts import TextArtifact
from griptape.utils.decorators import activity
from schema import Schema, Literal

class StockPriceTool(BaseTool):
    @activity(
        config={"description": "Get current stock price"},
        schema=Schema({Literal("symbol"): str})
    )
    def get_price(self, params):
        symbol = params["values"]["symbol"]
        price = fetch_stock_price(symbol)
        return TextArtifact(f"{symbol}: ${price}")

Memory & Conversation

from griptape.structures import Agent
from griptape.memory.structure import ConversationMemory
from griptape.memory.task.storage import LocalDiskTaskMemoryDriver

# Persistent conversation memory
agent = Agent(
    conversation_memory=ConversationMemory(
        driver=LocalDiskTaskMemoryDriver(file_path="memory.json")
    )
)

agent.run("My name is Alice and I work on ML pipelines")
# Later...
agent.run("What do you know about me?")
# -> "You're Alice, you work on ML pipelines"

Multi-LLM Support

from griptape.drivers import AnthropicPromptDriver, OpenAiChatPromptDriver

# Use Claude for reasoning
agent = Agent(
    prompt_driver=AnthropicPromptDriver(model="claude-sonnet-4-6")
)

# Or OpenAI
agent = Agent(
    prompt_driver=OpenAiChatPromptDriver(model="gpt-4o")
)

# Or local via Ollama
from griptape.drivers import OllamaPromptDriver
agent = Agent(
    prompt_driver=OllamaPromptDriver(model="llama3")
)

FAQ

Q: What is Griptape? A: Griptape is a modular Python framework with 2,500+ GitHub stars for building AI agents with composable structures (Agent, Pipeline, Workflow), pluggable tools, and Off-Prompt data processing for secure enterprise use.

Q: What makes Griptape different from LangChain or CrewAI? A: Griptape's key differentiator is Off-Prompt processing — sensitive data is processed locally and only summaries reach the LLM. It also has a cleaner modular architecture with three distinct structures, versus LangChain's sprawling API surface. CrewAI focuses on multi-agent roles; Griptape focuses on secure, composable pipelines.

Q: Is Griptape free? A: Yes, the framework is open-source under Apache-2.0. Griptape Cloud (managed hosting) is a paid service with a free tier.