Why connect local AI devices?
One computer may not have every model, tool, or hardware capability a team needs. A trusted-device network lets several approved machines contribute without treating unknown public providers as equivalent to devices owned by you, your family, or your organization.
The core idea is simple: the network owner approves participants, tasks are routed to suitable trusted devices, and outputs return to the requester. The useful questions are how membership is controlled, where the model runs, and what infrastructure can observe.
The main components
Users and requesters
A user gives an agent a task and selects the private network as its execution context.
Approved Neurons
A Neuron is a device contributing model or agent capability. Only devices approved by the network owner belong to a Local AGI Private Network.
End-to-end encrypted communication
Traffic between approved nodes is encrypted end to end. Intermediary infrastructure can route encrypted traffic but should not be able to read the task contents.
A relay for internet connectivity
Private devices are often behind routers and NAT. Local AGI is explicit that its Private Network is not pure peer-to-peer: internet traffic can pass through a relay operated by Local AGI. The relay acts as a blind pipe for end-to-end encrypted data; the AI execution remains on a trusted device.
End-to-end encryption protects content in transit. It does not make an approved endpoint trustworthy by itself. The network owner is responsible for approving devices and protecting them.
A task through the network
Select Private Network
The requester chooses the approved network instead of local-only or public execution.
Match capability
The task is directed toward a trusted node with the required model or agent capability.
Encrypt and relay
Traffic crosses the network as end-to-end encrypted data, using relay infrastructure when direct connectivity is not available.
Execute on the trusted device
The approved node runs the selected model or agent workload rather than handing the plaintext task to the relay.
Return and review
The result returns to the requester for human review and artifact handling.
Define the trust boundary
A useful private network policy answers:
- Who can approve or remove a device?
- Which models and agents can each device provide?
- What logs exist and who can access them?
- How are compromised, lost, or retired devices removed?
- Which workloads are permitted on the network?
- Which actions still require human approval?
“Private” should mean a documented boundary, not simply fewer users.
When several devices run one model
If a model is too large for one machine, several devices can each load a slice and run it together as a hive. A running hive can appear to the Private Network as one Neuron. Performance depends heavily on the connection between hive members, because intermediate model data must move between devices. See exactly how this differs from a multi-agent system in the AI Hive Mind guide.
Private network vs local-only vs public
| Execution | Best fit | Primary boundary |
|---|---|---|
| Local AI | Sensitive work that should stay on one machine. | Your device. |
| Private Network | Approved devices need to share models or compute. | Network membership and trusted endpoints. |
| Public Network | General tasks that benefit from broader available capacity. | Public provider policy and task suitability. |
For the broader tradeoff, see local vs cloud AI agents. For product details, visit the Local AGI Private Network feature page.
Build a network around devices you trust
Start locally, approve participants deliberately, and use private shared compute only where it improves the task.