DISPATCH runs entirely on your own machine — self‑hosted, no telemetry, no account. Your data, prompts and code never leave the box. Air‑gap friendly and offline‑capable.
DISPATCH is a runtime you clone and run — not a SaaS you log into. The graph, run engine, context store and every node execute as local processes. No hosted control plane, nothing phoning home. The only packets that leave the box are the model calls you wire to a cloud provider — bind to local models and there are none at all.
Everything runs inside the perimeter.
The run engine that drives Initialiser → Investigate → Plan → Create → Review is a process on your machine. Node state, the shared Context store, analytics and logs are all written to local disk. Wiring a graph, pressing Run, gating a Review checkpoint — none of it touches an external server.
Zero telemetry. By design.
Data residency, solved.
Regulated, sensitive, or just private work stays where it's generated. Source code, customer records and prompts are written to your disk — not to a vendor's logs.
Air-gap friendly.
Pair it with Local Models and a whole orchestration runs on a box that has never touched the internet. No external calls means no attack surface to the outside world.
Offline-capable.
Pull the network cable and the canvas keeps running. Local-bound nodes execute, edges animate, Review gates fire — the run never depended on a connection.
A chain runs inside your machine's perimeter. Every node tries to emit telemetry; every packet is stopped at the wall. Flip Air-gap to sever the network and confirm the run keeps flowing at 0 bytes egress.
Every node runs on this machine. Outbound telemetry never leaves the perimeter — egress holds at 0 bytes.
There is no onboarding flow and no cloud bootstrap. You clone the MIT runtime, point your nodes at the models you choose, and run. Connectivity is opt-in, per node.
git clone then start the runtime. It binds to localhost and serves the canvas to your browser — no account, no setup server, no callback.
Bind nodes to cloud providers if you want them, or to local models for zero egress. Outbound calls are exactly the ones you wire — nothing implicit.
State and logs land on your disk. Because the runtime is open source, you can read or instrument every byte that crosses the boundary — and confirm there isn't one.
Ollama, llama.cpp, LM Studio — fully local, alongside the cloud.
openName-tagged cursors, inline comments, live node-wiring — no cloud.
openShip a canvas to a GPU pod serving a 70B-class model.
open