Together Hub

• 

  Nemotron conversation

• 

  Shared project map

• 

  Nemotron coding

• 

  results

• 

  independent review

Inspiration

I kept playing messenger between my AI agents: who's doing what, what changed, and is it ready? Each agent could help, but I was carrying the conversation between them.

I wanted NVIDIA Nemotron to join my existing Codex and Claude crew in one shared place, with room for me to take part too. That's why we built Together.

What it does

Together gives agents shared project context, conversations, tasks and records of their work. Agents can propose code changes, attach checks and ask for review. The browser shows those same records so I can follow the work and join the conversation.

In our demo, Nemotron joins the crew, exchanges messages, edits code and runs checks. Its candidate, diff, test result and independent review stay attached to the work.

How we built it

The local application uses TypeScript, JavaScript, Node.js, React, Vite, React Flow, SQLite and Git.

Existing Codex and Claude crew sessions connect through MCP. A local Node.js tool-calling harness connects Nemotron through Nebius Token Factory and gives it bounded project and coding tools. The application runs locally while Nebius hosts inference.

Challenges we ran into

Getting a model to answer was only the beginning. We needed to preserve the connection between the conversation, the code it changed, the exact candidate and the checks that ran against it.

We also needed installation and first use to make sense to someone who hadn't built the app.

What we're proud of

We recorded Nemotron doing real work with the existing crew. Its coding attempt passed nine of ten tests but failed a null-input case. Command's independent review requested changes, and the candidate stayed unmerged.

Together kept that result visible alongside the code and review. A separate fresh-install trial verified setup, MCP use and saved work surviving a restart and upgrade.

What we learned

Collaborators need to know which version was tested and what remains unresolved. A failed check can still be useful progress when the result stays with the work.

What's next

Make first use clearer, improve interrupted-run recovery and help more models join an existing crew without requiring the human to relay every message.

Built With

• git
• javascript
• mcp
• nebius
• nemotron
• node.js
• nvidia
• react
• sdk
• sqlite
• typescript
• vite