models · local ← all features

Local Models.

Bind any node to a self‑hosted model — Ollama, llama.cpp, LM Studio, vLLM — and run fully offline, alongside or instead of OpenAI, Anthropic and Google.

In development OpenAI-compatible · zero egress

what it is

A node doesn't care where the weights live.

Every node carries its own model binding. Local Models points that binding at a server on your own machine — so Plan can stay on a hosted frontier model while Investigate, Create and Doc run on a quantised 8B from your laptop. Same graph, same Run button.

the binding

Provider → endpoint → model.

Pick local as the provider and DISPATCH asks for a base URL instead of an API key. It speaks the OpenAI /v1/chat/completions shape, so Ollama, llama.cpp, LM Studio and vLLM all drop in — no adapter to write.

alongside, not instead

Mix local and cloud on one chain.

Bindings are per-node, so a single run can fan across hardware. Keep sensitive context on a local model; reach for a hosted model only where you need the extra reasoning. The edges, status dots and analytics all behave identically.

Investigatelocal · llama3.1:8b

Plananthropic · claude-opus

Createlocal · qwen2.5-coder:14b

Doclocal · llama3.1:8b

why it matters

Private by construction.

Code, data and prompts never leave the machine. No vendor logs, no egress. Pull the network and the graph keeps running.

why it matters

Free iteration.

Re-run a fan-out of Parallel nodes a hundred times. The only cost is electricity — no per-token meter to watch.

why it matters

Air-gap ready.

No telemetry, no account. A whole orchestration can run on a box that has never touched the internet.

try it

Point a node at your own machine.

Pick a runtime — DISPATCH fills the endpoint, binds the Create node to a local model, and streams tokens straight off your hardware. Flip Offline to sever the cloud and watch the chain keep flowing.

local model · live serving locally