Sber builds out GigaChain without forking LangChain, betting on a compatibility package

Sber has explained how it builds an open source ecosystem around GigaChat and why it dropped its LangChain fork in favor of a separate compatibility package. The shift made AI agent development easier inside the company and made the stack clearer for external teams.

Why the approach changed

GigaChain appeared when Sber needed a single development tool for GigaChat, for both internal teams and external developers. Rather than write everything from scratch, the team decided to build on existing open source.

They chose LangChain as the base: it was the most visible framework for agent systems, it worked with many different models and it did not tie developers to a single vendor. For a large company, that choice meant less risk, a faster start and a clearer market of specialists.

At first, the team did go the fork route. The reason was practical: the prompt engineering era was only gaining momentum, and many prompts in LangChain were hard-coded in English. For GigaChat, which was better adapted to Russian, this created unnecessary constraints. Another problem was that full multilingual support and vendor integration were hard to push through quickly in the main project.

But the approach soon became too expensive to maintain.

"We had to pull in 100–200 changes that landed in the main project

literally every week."

When LangChain began changing rapidly and moving to a modular architecture, Sber abandoned the fork and built its own compatibility package. The recommended workflow now looks like this: a developer installs vanilla LangChain and adds the GigaChat integration package.

This proved more resilient: even after the package disappeared from the official LangChain documentation, almost nothing changed for users. In effect, the team kept compatibility with the market standard without getting stuck in endless synchronization of someone else's code.

What Sber is building

Around GigaChat, the team is developing not a single SDK but a whole open ecosystem. Its goal is not just to provide API access to the model but to shorten the path to a working AI agent. According to the team, a significant share of Sber's internal agent projects already relies on this stack, and external developers are using it more and more.

One signal of demand is the Python library for working with the GigaChat API: according to ClickPy, it ranks in the top 1.5% of the most downloaded packages on PyPI by monthly downloads.

• Python library for integrating with the GigaChat API
• compatibility package with LangChain instead of a custom fork
• open standalone agent GigaAgent
• documentation, integration examples and practical guides
• webinars, articles and conference demos

GigaAgent plays a separate role. The team did not build it as a lab demo project but refined it based on feedback from engineers and business. At dozens of conferences during 2025, developers showed the agent live and collected comments: where a REPL and tools are needed, where local execution is critical and where modularity matters more. Business, in turn, pushed toward applied scenarios: data analysis, preparing presentations and working with corporate sources.

What open source gives Sber

For Sber, the open approach solves several problems at once. First, hiring is easier: if the stack is built around LangChain and compatible libraries, the company does not need to look for people for a completely unique internal platform. Second, the barrier to moving existing agent solutions to GigaChat drops. The idea is simple: if a developer has already built an agent on a popular framework, it should be easier to adapt it to a Russian model without a full rewrite.

There is also a third effect: community help. The team gave an example with LlamaIndex: Sber originally did not support this framework because of limited resources, but an external developer later sent a pull request adding GigaChat support on top of an existing library. As a result, the ecosystem gained an integration with another popular stack at almost no internal cost.

The licensing policy remains as simple as possible: for its own projects, the team usually chooses MIT and closely watches the licenses and behavior of external maintainers.

For Sber, communication around open source is not a side activity but part of the product. The team answers on GitHub, maintains documentation, publishes articles, runs Telegram channels, holds webinars and appears at industry conferences. In 2025 alone, it gave more than 15 external talks. In this model, content exists not for PR but to shorten the time between a release, feedback and the next iteration of the product.

What this means

The GigaChain story shows that for corporate AI stacks today, compatibility with the market's main open source tools matters more than a custom fork at any cost. If Sber stays on this course, GigaChat could noticeably strengthen its position in Russian-language agent development and in enterprise scenarios, where speed of adoption and access to a familiar stack often matter more than technological exoticism.