Langflow

Langflow is an open-source framework for building agentic AI applications visually. It compiles flows down to real Python — no low-code tax — and plugs into every major model provider. As of April 2026 the project has 147,000+ GitHub stars, 8,700+ forks, and runs in production at hundreds of organisations.

I joined the core team through a path shaped by acquisitions: I was at DataStax from late 2022, DataStax acquired Langflow in April 2024, and IBM acquired DataStax in August 2025. I’m now one of the most active committers — 324+ commits across 73 active weeks, and 240+ merged PRs — working on the Developer and Operator Experience team.

What I’ve shipped

A non-exhaustive sampling of PRs I’ve authored and merged:

The Langflow SDK & Flow DevOps toolkit

PR #12245 introduced the langflow-sdk and a DevOps API toolkit, making it possible to manage flows, runs, and deployments programmatically without clicking through the UI. Followed up with #12679 wiring the SDK into the release workflow.

Model Context Protocol (MCP) integration

Langflow was an early adopter of Anthropic’s Model Context Protocol. I contributed many of the MCP plumbing PRs, including:

• #10965 — MCP + OAuth support
• #12662 — OAuth callback URL handling for MCP Composer
• #12715 — fixing auth error on MCP server restart
• #12622 — preserving optional field types through MCP Tools
• #12359 — supporting self-referential MCP JSON schemas
• #12746 — respecting renamed/deleted default folders across logins

Agent policy & ToolGuard

PR #12592 shipped a policies component via ToolGuard — so you can declaratively gate what tools an agent is allowed to call under what conditions. I also contribute upstream to ToolGuard itself.

ModelInput unification

PR #12025 was a cross-cutting refactor that unified the model-input selector across every component — the kind of boring, broadly-felt cleanup that pays dividends for years. It followed work on context-ID message history (#10319, #10183).

Document parsing with Docling

PR #9398 integrated IBM’s Docling for advanced document parsing in the File Component, and #12296 fixed its worker subprocess handling.

Model & embedding providers

• Astra DB Base Class (#10000)
• Ollama embeddings (#10356)
• watsonx embeddings + Batch Run
• Misc Gemini / Nvidia polish

What Langflow taught me

Working on a project this visible, with this many users, has forced habits I didn’t have in smaller codebases:

1. Big refactors only land if you split them into a boring PR and an interesting PR. The interesting one wouldn’t have gone in without the boring one first. The ModelInput unification is a good example.

2. Authentication is a systems problem, not a code problem. The volume of MCP auth bugs I’ve fixed (swap, restart, OAuth callback, folder rename) comes from the fact that a reasonable caller exists for every code path — you just don’t see them until production traffic hits.

3. “Backwards compatible” is usually cheaper than you think. When a model-provider SDK breaks its API surface every minor release, the cheapest thing you can ship is a thin adapter that your future self will bless you for.

4. The best open-source DX isn’t documentation — it’s running pip install and having the thing work on the first try, on Windows, with Python 3.11. A lot of my merged PRs are about that.