Compare
AI Stays Local and exo
Open-source software that connects your devices into an AI cluster.
When exo is the better fit
- You want to run one model across several of your own machines today.
- You have Apple Silicon machines to connect and want to use them together now.
- You want an open-source project you can study and modify.
Side by side
exo as its own public pages describe it on 2026-09-21, and AI Stays Local at the maturity its product data records. Where a source does not say, the table says so rather than guessing.
| Dimension | exo | AI Stays Local |
|---|---|---|
| Interface | An API compatible with OpenAI Chat Completions, Claude Messages, OpenAI Responses and Ollama. | Command line and an OpenAI-compatible local API, in private beta for selected testers. A desktop app is planned and not built. |
| Runs models itself | Yes, split across devices with tensor parallelism and topology-aware placement. | Yes. Inference sits behind an engine adapter; the recorded profiles run on upstream engines under their own licences. |
| Documents and retrieval | Not stated in the source checked | Folders you grant, lexical retrieval, answers that cite file and location, and refusal when the documents do not contain the answer. |
| Hardware planning | Placement takes device resources and the network into account, with automatic device discovery. | Evaluates the machine and recommends a model profile (private beta). Each model states execution, validation, usability, hardware and availability separately. |
| More than one machine | Yes. This is what it is for. | Planned. Multi-node execution is designed and not implemented. |
| Platforms | macOS with MLX, including RDMA over Thunderbolt 5, and Linux on CPU. | Command line on Linux and macOS for selected testers. Measurements published from one Linux ARM64 machine. |
| Licence | Apache-2.0. | Commercial product. A Free edition is defined for local use; paid prices are provisional. The runtime is not open source. |
Where AI Stays Local puts its emphasis
- Multi-machine execution is what exo does today. In AI Stays Local it is planned, not implemented, and this site does not describe it otherwise.
- AI Stays Local is aimed first at one machine you own, with a recommended profile, published per-model evidence and cited document answers.
- In both projects, several machines do not become one shared memory pool: a model split across machines moves data over the network on every token.
Sources, checked 2026-09-21: exo repository README. Descriptions of other projects are summaries of their own public pages on that date and may since have changed. No statement here is a claim about another project's security.