Julep is an open-source platform for building durable, composable AI agents — workflows that can run for hours or days, survive crashes, resume where they left off, and explain every step they took. It has earned about 6,600 stars on GitHub, is Apache 2.0 licensed, and is actively developed in 2026.
Most agent frameworks keep a workflow in one process’s memory: if it crashes, the work is gone. Julep instead treats agent workflows as durable programs. You compose tasks from steps — LLM calls, tool use, web search, conditionals, loops, parallelism and human-in-the-loop waits — and the platform persists state continuously, retries failures safely, and keeps an inspectable execution history. Agents maintain long-term memory and user sessions across conversations, and can schedule recurring work. SDKs for Python and Node.js let you define workflows as code, while built-in integrations cover common tools and document/RAG search over your own data. Executions can also be triggered by events and webhooks, so agents react to the outside world rather than only to chat messages. You can self-host the whole stack, or use Julep Cloud, the managed offering with usage-based pricing for teams that prefer not to run the infrastructure.
Freemium. The Julep platform is open source (Apache 2.0) and free to self-host; Julep Cloud provides managed hosting with paid, usage-based plans.
Developers building production agents and automations — research pipelines, support workflows, data jobs — where executions are long, stateful and too important to lose when a process restarts.