Multi‑User Group Chat Mode for LibreChat 🤝💬

[ghost]

Hi everyone,

first of all – thanks a lot for all the great work you’re putting into LibreChat. 💖 It’s an amazing project and we truly appreciate the flexibility and multi‑model integration it offers. 🤗✨

I’d like to share an idea as an impulse for future discussion:

Would it be possible to introduce a multi‑user group chat mode 🤝💬 where several authenticated users can join the same conversation with the assistant at the same time?

Potential benefits & use cases:

🧠 Team brainstorming sessions with AI in real‑time
🤝 Collaborative problem‑solving in small workgroups
💼 Joint customer support handling with AI assistance
🎓 Educational settings where a group of students interacts with the model together

I’m aware that such a feature might require significant changes 🛠️🥵 (real‑time backend, shared conversation storage, session synchronization) – but it could open up entirely new scenarios for LibreChat beyond single‑user interactions.

Question to the community :

Do you see interest or potential in this feature? Could we maybe collect feedback 📊 or run a small poll 🗳️ to understand whether this is worth considering for a future roadmap?

Thanks for listening 🙏 – I’d be happy to discuss requirements or contribute ideas for implementation.

Best regards,

Felix 😽

[kriskarkoski]

This is something we would be interested in. Beyond team collaboration we use Librechat for some support chats so it would be ideal to be able to handoff an existing chat to another support agent to continue with existing context, fork, etc.

[sempervictus]

Given the agent marketplace aspect already being in-place, collaborative interactions between N-users and N-agents seems like the next logical step vis-à-vis force multiplication. Ideally it would include letting each user bring their own keyed agents into the mix which they can direct to assist during the discussion (thinking "credentials for MCP client to a system other users/agents will be accessing with their unique authorization/authentication path to mutually enrich the conversation).

For support orgs and such this would be very handy as it would allow engineers and agents to reason-out a mutually acceptable/autonomically-informed response prior to having a "bridge agent" to their Slack/Teams/Discord respond to the user correctly in that medium. For commercial and non-profit entities, this is key to retaining the decision-making process data relevant to their regulatory constraints (a self-hosted librechat/openwebui/etc with internal inferencing and MCP back-ends addresses a LOT of GRC concerns).

On the practical side of "how do we manage turns?" to avoid every message causing a giant context to be reprocessed by a dozen agents reading it:

1. @ of the agent explicitly prompting it to issue completions for the next step appending all intermediate conversation in the channel/room (without threads) to its existing KV cache and focusing-down on the answer in the @ request
2. Thread a conversation from the main channel/room with relevant participants (human or not) performing "real-time" inference on the chat stream in that channel
3. Implement throttling on agents for request and data volume (along with basic loop defense) to avoid runaway scenarios where they hallucinogenically prompt each other into conversation eventually bankrupting the owner and flooding stacks with nonsense.

This would also lead to the ability to construct conversational task chains in some basic anchored syntax for @agent1 please research subect X and then ask @agent2 and @agent3 to post-process and validate the findings for @agent4 to produce a ship a report to @user1 before morning with conditionals and so on. Most humans aren't prompt engineers but they can figure out how to "converse" with properly prompted agents to do small/feasible things in logical series'

[adriangonzalezspectec]

We'd be very interested in something like this. Currently have an agent we use for iterating on requirements and would need to be able to open it to certain people in certain departments (Commercial, Product, Engineering, Support, etc) to give their input.

I understand that controlling who is currently talking to the agent could be a challenge, so perhaps the possibility to define per conversation:

• an owner
• assign to a user (and block any other from using it)
• owner can assign at any time to anyone or claim it back
• whoever is currently assigned can assign it to anyone else

[benhadad]

How about using swimlanes? Each user in the conversation could have their own lane, similar to how different chatbots are currently displayed. Each post would act as its own entry point, much like a forum thread.

A key feature would be allowing the user to decide what context is sent to the AI with each prompt:
* **Selective Context:** Everything up to the specific comment they are addressing.
* **Full Context:** Everything up to the most recent comment.

There could then be a **"Main Stream"** that uses predefined rules to summarize the various thoughts and combine them into a cohesive output.

[jjhidalgar-celonis]

this would be awesome