Designed to know less.
Most AI tools ask for more: more context, more access, more of your code in someone else's cloud. kn0t is built the other way round. Coordination needs to know who is working where, not what the code says, so the product is designed around the smallest set of facts that catches a collision.
This page is the technical contract: what kn0t sees, where verdicts happen, how it fails, and what is deliberately impossible. Every claim here is also on the product's own surfaces, so it can be checked rather than believed.
What leaves your machine, exactly.
Working solo, no work content at all: the Map, edit verdicts, the transcript record, and its search index are local files on your machine. On a team, your machine shares coordination metadata and the agreed contract history: who is working, in which repo, which file paths are claimed, and what was agreed. File contents, diffs, prompts, and transcripts are never sent. If it would show what you wrote, it does not go up.
kn0t is not a proxy.
kn0t sits beside your agents, never between an agent and its vendor. There is no traffic interception, no local root certificate, no forked CLI, and no patched vendor binary: your prompts and code travel exactly the paths they travelled before kn0t was installed. Integration rides the hook and plugin surfaces each supported tool documents, and can be unwired from the app in one step.
Verdicts are local, and the engine is deterministic.
A tangle is created when an observed edit overlaps an active claim: kn0t does not guess a future conflict from a prompt, and no LLM call decides a verdict. Detection runs on your machine from live claims, symbols, and dependencies; the edit path never waits on kn0t's service. Agents and people author every contract. kn0t routes proposals, records consent, and clears the held retry.
Failure is honest.
If the shared service is unavailable, the local edge keeps coordinating and enforcing its current rules, and cached team policy stays active: an empty or uncertain feed never means disarm. If the local edge itself is unavailable, the supported hook exits within its budget and the edit proceeds. The worst kn0t can do to you is nothing, and that is a design decision, not an accident.
Authority stays human.
kn0t identifies where work crosses and who has standing to decide; it does not invent the technical contract. The affected agents and people propose it, matching text records consent, and a genuine judgement call moves to an authorised person. That boundary holds even when kn0t starts a coordination-only turn in an idle session: an unattended turn can propose and counter, and can never ratify another. Agreement completes with a person, or with a session a person is driving.
Posture, not surveillance.
A head of engineering can see whether enrolled seats run the published policy and whether coordination work is unresolved. That never requires a feed of prompts, transcripts, or per-person activity: posture describes the control, not the person. Policy fingerprints show governed, attention, and changed seats, and content or command matching happens on the seat, with matched text redacted from anything that leaves it.
Data-loss detection, at the edge.
Today's guardrails already check paths, written content, and commands against deterministic rules, on the machine. The next module extends that discipline from patterns to meaning: reading what an agent is about to write, and what a prompt is about to say, for sensitive content such as credentials, keys, and customer data.
Classification will run on the seat's own hardware, on Apple's on-device foundation model, so your code and prompts are never transmitted for analysis: not to kn0t, and not to anyone. Turn the network off and it still works. Rules are centrally managed; evidence is never centralized: an org sees the class, the confidence, and the rule that fired, never the content. The first release is advisory, warning and recording, while your deterministic rules keep deny and ask. There is no cloud fallback, deliberately.
Agent speed is already here. The next engineering advantage is shared context, explicit contracts, and a neutral layer that helps every tool work from the same live reality, while knowing as little as that requires.