Local-first does not mean local-only. It means the user-visible control plane, critical policy decisions, durable project state, and evidence remain under user-governed control. Remote models or services may still be used when their capability is justified and the data flow is explicit. This adapts local-first software principles to intelligent execution. Source: Local-first research
Key takeaways
- Place data and authority before choosing a model.
- Local policy can govern remote inference.
- Offline continuity requires bounded local capabilities and deferred synchronization.
Placement decision matrix
| Factor | Prefer local | Prefer remote |
|---|---|---|
| Privacy | Highly sensitive source or prohibited egress | Approved data class with explicit transfer |
| Latency | Interactive control or device response | Batch task tolerant of network delay |
| Capability | Local model meets quality threshold | Specialist or frontier capability is necessary |
| Cost | Available local capacity and predictable use | Elastic or infrequent workload |
| Availability | Offline or degraded-network requirement | Reliable network and managed service objective |
Hybrid model routing
- Classify the objective and data.
- Apply prohibited-data and region rules.
- Evaluate local model sufficiency.
- Minimize and transform context before remote transfer.
- Record provider, model, policy decision, and returned artifact.
- Keep approval and final side-effect control in the user-governed plane.
Offline continuity
An offline-capable MIR can preserve task queues, local tools, project memory, policy, and evidence even when remote inference is unavailable. It should expose degraded capability rather than silently changing behavior. Synchronization needs conflict handling, provenance, and clear retention rules.
User-governed control
- Inspect context before it leaves the endpoint.
- Review which model and connector will receive data.
- Approve high-impact actions locally.
- Correct or delete memory without negotiating with a hidden agent state.
- Export evidence in a durable, portable format.
ONNX Runtime is one example of an inference layer that can support local execution, but MIR adds policy, tools, state, and evidence above it. Source: ONNX Runtime
Source record
References
Martin Kleppmann et al.. Ink & Switch. Published 2019; last reviewed 2026-06-20 UTC. Research essay.
National Institute of Standards and Technology. NIST. Published 2024-07-26; last reviewed 2026-06-20 UTC. Government profile.
- ONNX Runtime Documentation Primary source
Microsoft and ONNX Runtime contributors. ONNX Runtime. Published Current documentation; last reviewed 2026-06-20 UTC. Official documentation.