Griptape

Everything runs inside a structure. An Agent is a structure with one prompt task and optional tools; a Pipeline runs tasks in list order; a Workflow is a directed acyclic graph in which tasks name their parents or children and run in parallel as soon as their inputs are ready. Tasks read shared context such as parent_output or parents_output_text through Jinja templates.

Multi-agent setups are built by nesting structures. A StructureRunTask (inside a Pipeline or Workflow) or a StructureRunTool (given to an agent) runs another structure through a Structure Run Driver, which can be local or point to a structure deployed on Griptape Cloud. There is no manager agent that plans assignments; the developer lays out the task graph.

Conversation Memory keeps chat history and can be persisted through drivers (local file, Redis, DynamoDB, Griptape Cloud). Task Memory holds large or sensitive tool results off the prompt and lets later tool calls query them, and Meta Memory adds extra metadata to the model context.

unknown: the framework docs and the repository's docs folder were searched for approval, human-in-the-loop, pause and interrupt. Only generic on_before_run / on_after_run task hooks and event listener drivers are documented; no approval step or pause-and-resume mechanism for a person was found.

pip install "griptape[all]" -U

from griptape.drivers.structure_run.local import LocalStructureRunDriver
from griptape.rules import Rule
from griptape.structures import Agent, Pipeline
from griptape.tasks import StructureRunTask

# The default prompt driver uses OpenAI: export OPENAI_API_KEY first.
def researcher() -> Agent:
    return Agent(rules=[Rule("List three plain facts about the topic you are given.")])

def editor() -> Agent:
    return Agent(rules=[Rule("Turn notes into one short newsletter paragraph.")])

team = Pipeline(tasks=[
    StructureRunTask(structure_run_driver=LocalStructureRunDriver(create_structure=researcher)),
    StructureRunTask(("Rewrite these notes: {{ parent_output }}",),
                     structure_run_driver=LocalStructureRunDriver(create_structure=editor)),
])
team.run("Honeybees")
print(team.output.value)

• Requires Python 3.10 or newer.
• The quickstart uses OpenAI and expects OPENAI_API_KEY; other providers need their driver extra, e.g. griptape[drivers-prompt-anthropic] or griptape[drivers-prompt-ollama].
• uv add griptape / pip install griptape installs only core dependencies; many drivers fail to import until their extra (or [all]) is installed.
• Per the migration guide, MCPTool now needs the MCP Python SDK 2.x (install griptape/tools/mcp/requirements.txt), and its WebSocket transport was removed in favour of streamable_http.
• Templates now render in a Jinja2 sandbox, so private or dunder attributes in task inputs resolve to undefined.

Yes, as a client: MCPTool connects an agent to an MCP server over stdio or streamable HTTP and needs the MCP Python SDK 2.x. A2A and AG-UI support were not found.

The framework is Apache-2.0 licensed. The same company runs Griptape Cloud for hosting structures; no public pricing page was found, so its pricing is not stated here.

A Pipeline runs tasks one after another. A Workflow is a directed acyclic graph where tasks with finished parents run in parallel and can read all parent outputs.

Each agent is its own structure. A StructureRunTask in a Pipeline or Workflow, or a StructureRunTool given to an agent, runs another structure locally or on Griptape Cloud.

Python 3.10 or newer.