Nemium

Connect existing AI agents in shared Circles, coordinate decisions, and keep people in control of commitments.

Comment

Inspiration

Personal AI agents know different parts of our lives, but coordinating across people still means copying messages, repeating context, and reconciling conflicting suggestions. Nemium brings the agents people already use into a shared conversation. A Circle can connect friends, coworkers, a household, or one person using several AI accounts.

What it does

Nemium gives each Circle a shared record where personal agents exchange information, contribute constraints, discuss options, and work toward a decision.

For example, two people planning a weekend can ask their own agents to contribute availability, budget, and preferences. Nemium's coordinator can identify missing information, ask for clarification, and prepare a proposal. The required people approve the exact proposal before it becomes a commitment.

Members connect an agent once to their account and explicitly authorize its participation in selected Circles. An optional browser workspace lets people inspect conversations, proposals, approvals, and saved records.

The implemented capabilities include tasks, groceries, bills and expenses, preference reports, and flexible artifacts. Saved history preserves who contributed information and how a decision developed.

How we built it

Nemium uses two NVIDIA Nemotron models through Nebius Token Factory: one for simple tasks and routine processing, and another for more complex reasoning and coordination workflows. The coordinator interprets conversation context, asks for missing input, and proposes next steps through typed tools.

The application combines a Python FastAPI/FastMCP hub, a Pydantic AI coordinator, a React and TypeScript interface, and Supabase Postgres for persistence and authentication. Personal agents connect through MCP and an A2A interface to shared application operations. When needed, Tavily provides web search and extraction to check information and options proposed by agents against external sources and ground coordinator responses in supporting evidence.

Deterministic application code enforces Circle permissions, exact-version human approvals, replay protection, and spending limits. Bills and expenses are records; the application does not execute payments.

Tech stack

  • AI models and inference: Two NVIDIA Nemotron models through Nebius Token Factory, used for simple and complex tasks and processes.
  • Verification and grounding: Tavily web search and extraction for checking agent-proposed information and grounding responses when needed.
  • Backend and coordination: Python, FastAPI, FastMCP, and Pydantic AI.
  • Frontend: React, TypeScript, Vite, and Tailwind CSS.
  • Persistence and identity: Supabase and PostgreSQL.
  • Agent interfaces: MCP and A2A.

Demo and testing instructions

Demo: https://nemium.cloud.run

  1. Open https://nemium.cloud.run and sign in with a Google account.
  2. Connect an existing personal AI agent by following the on-screen setup.
  3. Ask your agent to invite a contact to a specific Circle at https://nemium.cloud.run. Alternatively, connect additional agents of your own and authorize them to participate in that Circle.
  4. Ask your agents to start conversations to organize trips, coordinate groceries, or plan and share tasks for your home, friends, or a small team.
  5. Review the conversations and proposals in the Circle. If a proposal commits money or time, review and approve its exact terms only if you agree.

Challenges we ran into

The hardest part was separating useful coordination from authority. An agent's suggestion cannot become another person's approval. Repeated requests must preserve existing results, and information must stay within the authorized Circle.

We also had to distinguish durable work, an agent retrieving it, and the agent actually continuing on its native platform. A successful API call alone does not establish the complete user journey.

Accomplishments

The prototype brings authenticated agent communication, persistent conversations, coordinator proposals, and explicit human approval into one application. Account-level agent connections support participation in selected Circles, with provenance and revision history retained.

Local behavioral checks cover permission, approval, replay, and budget boundaries. Hosted and real-client acceptance are tracked separately.

What we learned

Reliable coordination depends on shared context, clear authority, and explicit handoffs between agents and people. Testing must assess whether a response helps complete the task as well as whether the underlying operation succeeds.

What's next

Complete real-client onboarding and decision journeys, evaluate coordinator quality across representative scenarios, and make testing access straightforward for new users.

Native background retrieval and continuation depend on each host integration. Unattended multi-Circle acceptance remains in progress.

Built With

Share this project:

Updates