Docling is a powerful document processing library. It handles PDF parsing, layout analysis, table extraction, OCR, and multi-format export. It's the kind of specialized AI pipeline that does one set of tasks very well. But scaling it in production means running docling-serve, which brings its own set of architectural constraints: async task state pinned to a single process and a resource intensive scaling path.

Stepflow is a general purpose AI workflow system designed to solve exactly this class of problem. Its orchestration architecture provides resilient distributed execution state, HTTP SSE-aware load-balanced routing to worker pools, persistent result storage, and full observability out of the box.

After initial success with scaling Langflow throughput, we started looking for other high-value integration projects. After looking through docling throughput as part of some day-job work with OpenRAG, we started to wonder: how hard would it be to take a sophisticated AI pipeline like docling and run it entirely on Stepflow?

The answer was a lot simpler than we though: some quick dev work shaping requests and responses for API parity via a simple proxy and setting up a basic Stepflow flow to define the workflow. This post walks through how Stepflow's architecture made that speed possible and what it means for scaling AI related pipeline tasks like document processing in production. We also introduced some enhanced, docling-specific observability facilitated by this architecture for which any administrator who has had to wrestle with AI workflow issues like token burn will find immediate value.

In our first Langflow on Stepflow post, we showed how to convert and execute Langflow flows on Stepflow. That demo required building the Stepflow binary from source, generating configuration files, and wiring everything together manually. It worked, but it wasn't exactly a one-liner.

Today, the stepflow-langflow CLI makes this dramatically simpler, thanks partly to the new stepflow-orchestrator package which includes the necessary binary to run locally. A single run command handles conversion, orchestrator startup, and execution — all in one step.

Note: This post uses the legacy $from syntax. As of December 2025, Stepflow uses the new syntax with $step, $input, and $variable. See the latest documentation for current examples.

Stepflow was designed to provide a common runtime for GenAI workflow frameworks. We've been working on a proof-of-concept converting flows from Langflow to Stepflow and executing them using Stepflow. This post demonstrates the integration and provides implementation details.

Note: This post uses the legacy $from syntax. As of December 2025, Stepflow uses the new syntax with $step, $input, and $variable. See the latest documentation for current examples.

In the rapidly evolving landscape of AI applications, the ability to orchestrate complex workflows that combine language models with practical tools has become essential. Today, we'll explore how Stepflow, a modern open source workflow orchestration engine, seamlessly integrates LangChain's AI capabilities with common tools via the Model Context Protocol (MCP) to create a powerful research assistant.

While invoking MCP tools from within an agent is nothing new, in this post we are building a practical, modular "flow" on top of Stepflow's declarative workflow engine. This system will generate research questions, analyze text, create structured notes, and save everything to an organized file structure. All this will be orchestrated through a declarative YAML workflow that's easy to understand and modify.

We're excited to announce the initial release of Stepflow, an open-source protocol and runtime designed to make building, executing, and scaling GenAI workflows simple, secure, and portable. Whether you're prototyping locally or deploying to production, Stepflow provides the foundation for reliable AI workflow execution.