julep
Julep — durable, composable AI agents. Flows that crash and resume, retry safely, and explain every step.
What is julep?
Julep — durable, composable AI agents. Flows that crash and resume, retry safely, and explain every step.
Overview of julep
What it is
Julep is an open-source framework for building durable, composable AI agents. It structures agents as dataflows rather than ad-hoc loops, allowing flows to crash and resume, retry safely, and explain every step. The primary authoring surface is a define-by-construction @flow decorator that compiles ordinary Python code into a frozen wire-format IR. The pure core is dependency-free, with an optional Temporal layer for production durability. It is licensed under Apache-2.0 and is currently available as a release candidate (version 3).
Key features
- Flows can crash and resume, retry safely, and explain every step through a derived projection.
- Agents are defined using
@flow,@tool,@pure, andReasonercomponents that compile to a frozen IR. - Includes a CLI (
julep) for discovering, running, linting, testing, and deploying agents across a module. - Supports MCP (Model Context Protocol) tools with snapshot-backed schemas and safety checks.
- Production features include a self-hosted control plane with encrypted vault, per-run credentials, and MCP preflight validation.
Use cases
- Building durable AI workflows that require safe retries and crash recovery.
- Composing multiple agents into a cross-agent DAG for complex tasks.
- Developing and testing agents locally with dry-run and fake reasoners.
- Deploying agents to production with Temporal for durability and observability.
Getting started
Install the base package with pip (requires --pre flag for the release candidate): bash pip install --pre julep For a full quickstart example, see the repository README.
License and pricing
The project is licensed under Apache-2.0 and is open source. Pricing is not stated.
Based on the project's public README and repository metadata.