System · the mesh, it wires

Nexus

The layer it all runs on.

Nexus is the self-hosted platform that builds, wires and runs the systems and their stores on your own infrastructure, in the EU. An assistant operates it day to day through an audited control surface; anything with real-world impact consumes a one-time human grant. It is the ground the product stands on, not a feature of it.

The infrastructure

How it flows

Systems + stores
cortex · lens · atlas · …
Nexus mesh
Self-hosted · EU · wired
  • Wires every system and store together over a private mesh.
  • Self-hosted: your data never leaves your infrastructure.
  • Automatic TLS, zero-downtime deploys, and a one-time human grant for every external action.

How it fits

Runs and connects Cortex, Lens, Atlas, Thalamus, Larynx, Pharynx and the stores.

Under the hood

An assistant at the controls. A human on the grants.

Nexus is our own deployment platform: one governed engine that builds from git, fronts every app with automatic TLS, wires credentials between systems, and runs the whole fleet across a small cluster of machines. The unusual part is who operates it. An assistant does the day-to-day through the same audited control surface a human uses, and anything with real-world impact stays behind a one-time human grant.

The console · shown with demo data

The nexus canvas with demo data: five apps as nodes wired to a public router, a shared-credential group, and a deploy soaking in the live queue

Apps are nodes on a canvas. A cable to the router exposes one publicly, a dashed hull is a shared-credential group, and the queue at the bottom is a release soaking before cutover. Same fleet, phone-sized, on the right.

A deploy, end to end

git push
or one sentence in chat
Build
isolated image
Health gate
probe must pass
Dark-bake soak
0% of traffic
Cutover
transactional · zero-downtime
Rollback
the old release stays

A new release builds in isolation, must pass its health probe, then runs dark for a soak window while production traffic stays on the old one. Only a release that stays healthy is cut over; one that fails never replaces what is live. Stateful apps and stores deploy stop-then-start with automatic rollback.

Who may do what

Assistant
audited surface
Read
logs · status · stats
Change
env · limits · wiring
Deploy · expose · delete
external
Operator grant
one-time · it ships
∅ Refused
no grant, no action

Reads run instantly. Config changes ask once. Deploys, exposures and deletions are external actions: each consumes a grant issued by a human operator, and the assistant cannot issue its own. The refusal is enforced in the engine, not by a prompt asking nicely.

One session from our own fleet · names changed

09:12 · operator

“Sweep the runtime logs of all eight services and find anything noisy.”

09:13 · assistant

Reads eight log tails in one pass, secrets already redacted. Finds a worker throwing a socket timeout every five seconds on empty queues.

09:15 · assistant

Patches the queue timeout, adds a regression test, commits.

09:15 · assistant

Requests a deploy. That is an external action, so it needs a grant.

09:16 · operator

Issues a one-time deploy grant for the worker.

09:16 · nexus

Build → health gate → dark-bake soak → cutover. Rollback armed the whole way.

09:21 · nexus

Worker live. Logs quiet.

Nine minutes from question to quiet logs. The only human actions were the question and the grant.

3
machines paired into one cluster, including a GPU host
29
apps and stores of ours run on it today
1
one-time human grant consumed by every deploy
~30%
of a PaaS, on purpose: the slice a small fleet actually uses
Environment changes that used to be an afternoon of clicking are one sentence in chat; secrets can be minted server-side, so the assistant orders a credential but never sees its value.
Every log the assistant reads is redacted first, and build logs survive failures in the shared cluster store: a deploy that died at 03:00 is diagnosed from any machine at 09:00.
Backing stores come from recipes (Postgres with pgvector, Qdrant, Neo4j, Redis, MinIO); consumers get the protocol-correct connection env injected on their next deploy.
Machines pair into the cluster with a code and share one state: every box can answer for the whole fleet, and each app pins to the machine that fits it, GPU where needed.

The rest of the constellation

This is one part of a system built to keep one promise: every answer is grounded and permission-aware, or it doesn’t come at all.