NVIDIA
Offline Protocol and NVIDIA are building for AI that runs where the work happens. As part of NVIDIA Inception, we build the coordination layer that lets on-device agents discover, authenticate, and call each other across devices, with no cloud round trip and no broker in the middle.
AI that coordinates where the work happens
Inference has moved onto phones, robots, vehicles, and edge computers. Coordination has not: two agents on adjacent devices still talk through a data center that may be seconds away or unreachable. Offline Protocol is the layer that closes that gap, and as part of NVIDIA Inception we build it with NVIDIA’s developer tools, training, and technical guidance.
The result is a mesh of agents that find each other, verify each other, and share work directly, on the same devices where the models run.
What your agents can do
Your agents get a coordination layer that works with no network at all, and reaches further as more devices join.
- Service discovery lets an agent advertise a tool or a model and call it on another device, with no broker in the middle
- Agents discover peers up to 8 hops away over Bluetooth LE and Wi-Fi
- Requests are authenticated with OfflineID, the identity every device holds
How to get started
- Read how agents advertise and call capabilities over the mesh on the AI agents page.
- Connect your coding agent to the Offline Protocol MCP server to plan and scaffold the integration.
- To run agents across your own devices, talk to us about an evaluation.
NVIDIA Inception FAQ
Can AI agents work without the internet?
Yes. With service discovery, an on-device agent advertises a tool or a model, peers up to 8 hops away discover it over Bluetooth LE and Wi-Fi, and they call it request and response inside an OfflineID-authenticated MLS session, with no broker, server, or DNS between the devices. Read about offline AI agents.
How is service discovery different from MCP?
MCP standardises how a client talks to a tool server. Service discovery removes the server: the capability lives in the mesh and travels with the devices. The two coexist, so you can keep an MCP-shaped interface for your agents and run the mesh underneath it as the transport. How service discovery works.
How does an agent choose between several peers offering the same tool?
Every provider in range answers with its version and capabilities, and the agent compares them by hop distance and version before sending anything.
Can AI agents share state across devices while offline?
Yes. With edge sync, several devices edit the same document with no link between them, and when they meet the edits merge to the same result on every device. No server decides who wrote last. How edge sync works.
Can I watch it work before I build?
The live demo runs in your browser and shows the SDK choosing a path between devices. The AI agents page walks through an agent calling translate.v2 on a peer two hops away. Open the live demo.
