Repolode
Submit
julep-ai

julep-ai/julep

Python framework for building durable AI agents as composable dataflows that can crash, resume, retry safely, and explain each step.

View on GitHubWebsite
Stars
6,574
Forks
966
Open issues
4
Last commit
3 months ago

What is it?

Julep builds AI agents as composable, durable dataflows rather than ad-hoc loops. Flows can crash and resume, retry safely, and produce a derived projection that explains every step. The model can only call tools it was explicitly allowed to use.

The main authoring surface is a `@flow` decorator. Ordinary Python functions define graph steps, while registered tools, pure functions, reasoners (LLM calls), branches, fan-out, retries and timeouts compile to a frozen wire-format IR. The core package depends only on PyYAML, and the Temporal runtime is an optional extra.

The package includes a `julep` CLI for listing, linting, testing, running, tracing and deploying agents in a project, along with a `dry_run` mode that executes flows locally with in-memory tools and fake reasoners. Optional extras add Temporal or DBOS execution, MCP tool references, OpenTelemetry and Langfuse export, and a self-hosted FastAPI control plane.

Key features

  • 01Define agent workflows with a `@flow` decorator using ordinary Python functions.
  • 02Flows can crash and resume, and steps can be retried with timeouts.
  • 03Tools the model can call must be explicitly allowed; others are denied.
  • 04Local `dry_run` executes flows with in-memory tools and fake reasoners.
  • 05A `julep` CLI lists, lints, tests, runs, traces and deploys agents.
  • 06Optional extras support Temporal, DBOS, MCP, OpenTelemetry and Langfuse.

Use cases

  • Building LLM workflows that must survive process crashes and resume safely.
  • Restricting which tools an agent can invoke for compliance or safety reasons.
  • Tracing and auditing each step of a multi-step agent run.
  • Running agent pipelines in CI with lint and test commands.

Quick start

  1. 1

    Install the package (pre-release flag required)

    pip install --pre julep
  2. 2

    Save the README quickstart code as a Python script and run it; no API key is needed for the dry run

Things to know

  • Julep 3 is a release candidate, so the --pre flag is required and APIs may change before 3.0.0 final.
  • Durable runtime features need optional extras such as julep[temporal]; the base install only authors and compiles flows.

In collections

Open-source AI agent frameworks & appsOpenTelemetry & observability platforms

Topics

#agents#llm#workflow#durable-execution#python#temporal#observability#ai#ai-agents#ai-agents-framework

FAQ

What is julep?
Python framework for building durable AI agents as composable dataflows that can crash, resume, retry safely, and explain each step. Julep builds AI agents as composable, durable dataflows rather than ad-hoc loops. Flows can crash and resume, retry safely, and produce a derived projection that explains every step. The model can only call tools it was explicitly allowed to use.
Is julep free to use? What license does it have?
julep is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is julep still maintained?
Updated within 6 months. The last commit to julep was on Aug 6, 2026.
What can I use julep for?
Building LLM workflows that must survive process crashes and resume safely. Restricting which tools an agent can invoke for compliance or safety reasons. Tracing and auditing each step of a multi-step agent run. Running agent pipelines in CI with lint and test commands.
How do I install julep?
Install the package (pre-release flag required): pip install --pre julep Save the README quickstart code as a Python script and run it; no API key is needed for the dry run
Where can I get julep?
The source code and installation instructions are on GitHub: https://github.com/julep-ai/julep

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects