kitaru

Kitaru: durable execution for AI agents.

Kitaru provides primitives for making AI agent workflows persistent, replayable, and observable. Decorate your orchestration function with @flow and your work units with @checkpoint to get automatic durability.

Example

from kitaru import flow, checkpoint

@checkpoint
def fetch_data(url: str) -> str:
    return requests.get(url).text

@flow
def my_agent(url: str) -> str:
    data = fetch_data(url)
    return data.upper()

Current status:

• Implemented: @flow, @checkpoint, kitaru.log(), save(), load(), wait(), llm(), memory.configure/set/get/list/history/delete(), connect(), configure(), stack lifecycle helpers (list_stacks(), current_stack(), use_stack(), create_stack(), delete_stack()), model alias helpers via CLI (kitaru model register/list), KitaruClient execution/artifact/memory APIs (get/list/latest/logs/input/retry/resume/cancel/replay + artifacts + memories), and a typed Kitaru exception hierarchy with failure journaling (Execution.failure, CheckpointCall.attempts).
• Implemented: replay support (KitaruClient.executions.replay(...)).

The CLI also supports global runtime log-store configuration via kitaru log-store set/show/reset, stack lifecycle via kitaru stack list/current/use/create/delete, and execution lifecycle commands via kitaru executions get/list/logs/input/replay/retry/resume/cancel.