julep-ai/julep

What it solves

Julep addresses the fragility of AI agents built with ad-hoc loops, which often crash, lack transparency, and are difficult to manage in production. It provides a framework for building agents as durable, composable dataflows that can safely retry, resume after crashes, and provide a clear explanation of every execution step.

How it works

Instead of runtime loops, Julep uses a "define-by-construction" approach where agents are authored as @flow functions in Python. These flows are compiled into a frozen wire-format intermediate representation (IR) that describes the graph of steps.

Key components include:

• Registered Tools and Pures: Explicitly defined functions that the agent can call.
• Reasoners: LLM-powered components that process prompts and return structured data.
• Durable Execution: An optional Temporal layer allows flows to persist state, enabling them to survive crashes and resume exactly where they left off.
• Control Plane: A self-hosted FastAPI control plane manages deployments, secret vaults, and MCP (Model Context Protocol) safety checks.
• CLI: A terminal-native toolset for discovering, linting, testing, and deploying agents as a module.

Who it’s for

Developers building production-grade AI agents who require high reliability, strict tool-access control, and the ability to audit and trace complex agentic workflows.

Highlights

• Durable Dataflows: Flows can crash and resume without losing progress.
• Frozen IR: Compiles Python definitions into a stable format to prevent runtime drift.
• MCP Integration: Supports Model Context Protocol for tool references with snapshot-backed safety and preflight checks.
• Production-Ready Tooling: Includes a CLI for graph visualization, linting, and deployment to environments like staging or production.
• Secure Secret Management: Features a write-only operator vault and encrypted Temporal payloads for sensitive credentials.