# Privacy, air-gap, and agents

> Air-gapped and on-prem typed decisions with Laya vs hosted Jev agent patterns. Tool routing, MCP, and when open weights beat API keys.

## Answer

Agents still need a chat or planner model. Laya answers bounded questions locally so sensitive state never leaves the box for those steps.

When local Laya belongs in the agent loop. And how that differs from bolting Jev MCP tools onto a chat model.

## Why agents care about typed decisions

Tool-selection and policy gates are judgment calls, not creative writing. /guides/jev-with-ai-agents documents hosted Jev patterns (MCP, skills, middleware). The same architecture slots local Laya where API egress is forbidden or milliseconds matter. See /learn/jev-vs-llm-classification for why probabilities beat YES/NO parsing.

## Air-gap and regulated data

After you download Apache 2.0 weights, Laya inference can run without outbound calls, subject to your security review and supply-chain policies. Hosted Jev requires sending state to TypeSafe infrastructure. Use it when that trade is acceptable and contracted.

## Local loop example

The showcase snake demo (/showcase?demo=laya-mlx-lonely-mh) is a visceral example: many decisions per second on Apple Silicon with roughly a gigabyte-class memory footprint in community claims. Verify on your hardware.

## Composing with MCP

Nothing stops you from exposing a private HTTP or MCP tool that wraps local Laya predict() while Claude or Cursor still handle language. That mirrors Jev MCP servers documented in agent guides, but you own uptime and versioning.

## Limits for agent builders

Laya will not draft user-visible apologies, summarize threads, or write code. Keep generative models for language. Use Laya for which tool, which route, and whether to escalate. Same division of labor TypeSafe describes for Jev.

## In the directory

- [Laya vs Jev hub](https://jev.aitools.fyi/learn/laya-vs-jev)
- [Laya vs Jev compared](https://jev.aitools.fyi/learn/laya-vs-jev/compare)
- [System One primer](https://jev.aitools.fyi/learn/system-one)
- [What Jev is](https://jev.aitools.fyi/learn/jev-typesafe)
- [Benchmarks](https://jev.aitools.fyi/categories/benchmarks)
- [SDKs and clients](https://jev.aitools.fyi/categories/sdks)
- [@receptron/laya](https://jev.aitools.fyi/tools/receptron-laya)
- [ollaya](https://jev.aitools.fyi/tools/ollaya-dev-ollaya)
- [Laya-CoreML](https://jev.aitools.fyi/tools/mizorewww-laya-coreml)

## FAQ

### Is local Laya more private than Jev?

It can be, if weights and inference stay inside your trust boundary. Misconfigured logging or crash dumps can still leak state. Privacy is a system property, not a model license.

### Should I replace Jev MCP with Laya?

Only when self-host is a requirement. Many teams keep Jev MCP for IDE agents and add Laya only in deployed services.

## Sources

Directory copy checked 2026-09-26. Primary sources:

- [Hugging Face: convaiinnovations/laya (model card and benchmarks)](https://huggingface.co/convaiinnovations/laya?ref=jev.aitools.fyi)
- [GitHub: NandhaKishorM/laya (upstream SDK)](https://github.com/NandhaKishorM/laya?ref=jev.aitools.fyi)
- [PyPI: laya package](https://pypi.org/project/laya/?ref=jev.aitools.fyi)
- [TypeSafe docs: System One / Jev API](https://docs.typesafe.ai/?ref=jev.aitools.fyi)
- [Laya-MLX: BENCHMARKS.md (M3 Max, independent port)](https://github.com/mizorewww/laya-mlx/blob/main/BENCHMARKS.md?ref=jev.aitools.fyi)
- [Jev with AI agents hub (hosted patterns)](https://jev.aitools.fyi/guides/jev-with-ai-agents)

Canonical: https://jev.aitools.fyi/learn/laya-vs-jev/privacy-and-agents
Markdown: https://jev.aitools.fyi/learn/laya-vs-jev/privacy-and-agents.md