Zulip 12.0 Released

Zulip 12.0: Organized chat for distributed teams

We’re excited to announce the release of Zulip Server 12.0, containing hundreds of new features and bug fixes: end-to-end encryption for mobile push notifications, a major upgrade for Docker, configurable image previews, and much more! Almost 5,500 new commits have been merged across the project since the 11.0 release in August 2025.

Zulip is an organized team chat application ideal for both live and asynchronous communication. With Zulip, you own your data: it’s 100% open-source software, with easy migration between cloud hosting and self-hosting, plus a powerful API.

Zulip is developed in collaboration with our community. A total of 160 people contributed commits to Zulip since the 11.0 release, bringing the project to 1,680 code contributors. Zulip is remarkable for its number of major contributors, with 99 people who’ve contributed 100+ commits.

Huge thanks to everyone who’s contributed to Zulip over the last several months, whether by writing code and documentation, reporting bugs, suggesting features, translating, supporting us financially, participating in discussions in the Zulip development community, or just sharing ideas! We could not do this without the hundreds of people who help the project.

Project highlights

Today marks a release of the Zulip server and web application. We’d also like to share news and updates for the project as a whole since the 11.0 release last August:

Mobile app

Zulip now encrypts your push notifications end to end from the Zulip server to your mobile device. The encryption happens seamlessly when you use Zulip Server 12.0 with an up-to-date version of the Zulip mobile app for Android or iOS. We’re excited about how this upgrade makes Zulip’s security model simpler and more robust, offering self-hosted organizations complete control over their data.

We’ve been busy with lots of other improvements to the mobile app too, in eight mobile releases since the 11.0 release: easier navigation, managing your channel subscriptions, support for new Zulip features like inline images and channel folders, and other improvements too numerous to mention.

Next up, our mobile agenda includes making notifications on iOS more helpful, a “recent conversations” view like Zulip has on the web, and a long list of other features. As always, your feedback helps us choose what to build next; if there’s something you’re especially keen to see added or changed, please let us know.

AI tools are changing our development process

Over the past few months, AI has changed how we work.

Holding back the sea of AI slop

October through February was a challenging time for the project. This is always the season when many Google Summer of Code hopefuls first try their hand at contributing, and a time when the direct benefit to the project of reviewing outside contributions is lowest. But this year the floodgates opened (as they have for many OSS projects), and the attention required to identify good work in the sea of AI slop significantly impacted our ability to complete important projects. Of the hundreds of AI-generated PRs we received in this time, virtually none were merged.

I seriously considered banning LLM use for Zulip contributions. But our view is that contributors should be allowed to use modern tools in the service of producing great, reviewable work. AI-assisted work is of course subject to the same rigorous review processes we’ve always used for community contributions.

So we decided to invest in creating, refining, and enforcing a new AI use policy, which has the following key tenets:

End-to-end human responsibility for work and the communication around it. You always need to understand, test, and explain the changes you’re proposing to make, whether or not you used an LLM as part of your process to produce them. Clear and concise communication about points that actually require discussion. While we allow carefully edited AI-generated PR descriptions, we’ve had to ban AI-generated chat messages in the development community as too disruptive.

Manual enforcement of this policy has been rough, with far more PRs closed without review, stern warnings, and outright bans of repeat offenders than we’ve ever had to apply before. We expect that next fall, automation or other major changes will be required for the PR triage process to be manageable.

In the meantime, 2026 is the 11th consecutive year that Zulip is participating in Google Summer of Code — we’re looking forward to welcoming this year’s cohort soon!

Making AI tools useful for our development work

On a rare free weekend in January, I ran the experiment of using Claude Code myself, to see if I could steer it to produce good work for Zulip. The results were promising — enough for us to start investing in teaching Claude Code how to self-review its work, and how to produce PRs that are easy for maintainers to review. This has largely been an AI-supported process of digesting our contributor documentation into CLAUDE.md, and iterating when we see the model struggle.

As of February, maintainer review identified several issues in every AI-generated PR. So the best work we got out of it at that time was audits: security audits, as well as checking for bugs in our data import/export systems, or performance issues in key code paths. These use cases are structurally parallelizable search problems, where AI’s ability to scale was most fruitful.

But as I write this in late April, with the investment we’ve made to set up our repositories for AI-assisted work, our engineers regularly produce PRs with Claude Code that are basically ready to be merged. The starting point can be just a link to an issue or to a Zulip conversation. Claude Code reads the linked discussion and any further context it points to, and often puts together a reviewable PR with just a handful of iterations. Maintainers reviewing such AI-assisted work (myself included) have been finding the quality to be just as high as what our core team has historically produced without AI.

This new capability has allowed us to make improvements that previously wouldn’t have been worked on at all, and has enabled some new ways of working. Alya, our PM, used Claude Code to produce a 16-commit redesign of our integrations directory, which the area maintainer approved with no changes. The project took a lot of back-and-forth — visual changes are especially hard for the models to self-review — but enabled rapid iteration.

At this stage, AI technology provides us with a moderate speed-up of regular work and some helpful new workflows — not enough, perhaps, to merit so much attention. But this technology is improving at a remarkable rate, and seems on track to deeply alter how we develop software.

Try Zulip anonymously in a demo organization

As part of evaluating Zulip, you can now create a demo organization in Zulip Cloud without providing your email or any other identifying information. It’s a great way to explore the app, whether you’re considering Zulip Cloud or a self-hosted deployment. Demo organizations are automatically deleted after 30 days, but can be converted into a regular organization if desired.

Learning about Zulip

A revamped security overview describes Zulip’s security model and the details behind it. Secure by default: Your data is protected out of the box. Well-documented and easy to understand, so that you’re never caught by surprise. Flexible, so that you can configure Zulip according to your organization’s needs.