AI integrations

Zulip’s topics organize conversations within a channel, which is ideal for integrating with AI systems and collaborating with AI agents. With Zulip’s structure, it’s easy to prompt an AI with the appropriate context for what you want to accomplish.

You can connect your AI models of choice with Zulip using the interactive bots API, which makes it convenient to have AI models participate in conversations in whatever ways your organization finds most effective as the technology evolves.

Future Zulip releases will also contain built-in AI features, such as topic summarization. A major advantage of self-hosting your team chat system in the age of AI is that you maintain full control over your internal communications. It’s up to you how you allow third parties and AI models to process messages.

Zulip Server 10.x includes a beta topic summarization feature that is available for testing and experimentation. We appreciate any feedback on your experience, and what configuration options and additional features your organization would find useful.

Built-in AI features

Data privacy

Making sure customer data is protected is our highest priority. We don’t train LLMs on Zulip Cloud customer data, and we have no plans to do so.

We are committed to keeping Zulip 100% open-source, so the source code that defines how data is processed is available for third parties to review and audit.

General configurations

Self-hosted Zulip installations can choose whether to self-host their own AI models or use a third-party AI model API provider of their choice. Zulip’s AI integrations use the LiteLLM library, which makes it convenient to configure Zulip to use any popular AI model API provider.

  • Server settings: You can control costs using INPUT_COST_PER_GIGATOKEN, OUTPUT_COST_PER_GIGATOKEN, and MAX_PER_USER_MONTHLY_AI_COST settings, which let you set a monthly per-user AI usage budget with whatever pricing is appropriate for your selected model.

  • Organization settings: Administrators can specify who can use each AI feature that is enabled by the server. The permission can be assigned to any combination of roles, groups, and individual users.

  • Personal settings: Users who find AI features intrusive or distracting can hide them from the UI with a Hide AI features personal preference setting.

Topic summarization beta

Note

Topic summarization is not yet available in Zulip Cloud.

The Zulip server supports generating summaries of topics, with convenient options for doing so in the web/desktop application’s topic actions menus.

How it works

Warning

As with all features powered by LLMs, topic summaries may contain errors and hallucinations.

The topic summarization feature uses a Zulip-specific prompt with off-the-shelf third-party large language models.

When a user asks for a summary of a given topic, the Zulip server fetches recent messages in that conversation that are accessible to the acting user, and sends them to the AI model to generate a summary.

Emoji reactions, images, and uploaded files are currently not included in what is sent to the AI model, though some LLMs may have features that might follow links in content they are asked to summarize. (Note that Zulip’s permissions model for uploaded files will prevent the LLM from accessing them unless the files have been posted to a channel with the public access option enabled.)

Enabling topic summarization

Important

If you use a third-party AI platform for topic summarization, you are trusting the third party with the security and confidentiality of all the messages that are sent for summarization.

Enable topic summarization by configuring TOPIC_SUMMARIZATION_MODEL and related configuration settings in /etc/zulip/settings.py. Topic summarization and settings for controlling it will appear in the UI only if your server is configured to enable it.

Choosing a model

When modeling the pricing for a given model provider, you’ll primarily want to look at the cost per input token. Because useful summaries are short compared to the messages being summarized, more than 90% of tokens used in generating topic summaries end up being input tokens.

Our experience in early 2025 has been that midsize ~70B parameter models generate considerably more useful and accurate summaries than smaller ~8B parameter models.