Bring your own model, and keep it local

Most AI products answer by placing a call to an outside model: your question, and the passages that support it, are sent to a provider that runs the model and sends text back. Running locally keeps that work in the building. One command starts the Node Runtime with its own store, embedder and reranker, and the image carries those model artifacts, so once it is pulled the stack needs no network at all.

knaix local up

There is no account, no login, and no token. Any command addresses that node by name, the same way you would name a hosted node: knaix upload -n local ./docs uploads a folder into it, and knaix chat -n local asks it a question.

To get real answers, point the node at a model server already running on your machine. knaix local setup probes the ports Ollama, LM Studio, vLLM and llama-server listen on, lists the models each one actually hosts, and remembers your choice:

knaix local setup

You can also name the server directly with --model-url and -m instead of picking interactively. Either way the choice is remembered, so a later start is just knaix local up. The model runs on your machine, alongside the node, so the passages retrieved from your documents and the answer written from them stay on the same hardware.

Before you attach a model, generation falls back to a deterministic mock so you can see the pipeline work. Because the difference between a real answer and a placeholder matters, the mock says so at every layer: the answer is prefixed, a footer follows it, and JSON output reports the model as mock. Retrieval, reranking and citations are real in both cases.

Local mode is a single-machine runtime, not the hosted mesh. It runs one node on your own hardware and needs a container runtime to pull and run the image. It is the right tool for trying the platform end to end, and for keeping a workload entirely on one machine; it is not the multi-node tailnet, the managed provisioning, or the cross-node federation that the hosted tiers add. Answer quality tracks the model you bring: a small local model reasons like a small local model, and the bundled embedder is compact and strongest in English, so retrieval quality tracks that too.