a8 console

The failure it was built for

A gateway that lies with a 200.

Your gateway advertises a model. The machine behind it was rebooted, or reconfigured, or quietly loaded something else. The route still resolves. Every caller still gets HTTP 200. Nothing alerts, because from the outside nothing is broken — and you find out from an answer that is subtly worse than it should be.

a8 resolves each route to the machine actually behind it and asks that machine what it is serving. When the two disagree, it says so, and names which one it checked. Where a gateway cannot say where a route points, a8 tells you that too, instead of implying a check it never ran.

What it does

Find, plan, watch.

Discover

Probes conventional ports and identifies what answers — Ollama, llama.cpp, vLLM, SGLang, oMLX, mlx-lm, LM Studio, Foundry Local, FastFlowLM. On a mesh it reads the host list from your own client instead of scanning a range.

Approve a plan

Findings are a proposal, never a fact already applied. You edit it, you confirm it, and that becomes the thing reality is measured against. It is a plain file you can read.

Catch drift

Missing endpoints, unexpected ones, a model that changed underneath a route. Each finding says what was checked and how confident it is.

Start from nothing

No models yet? It installs the basics that suit your hardware, shows every command before running it, and confirms the port answers afterwards — because an exit code is not proof.

Control

Start, stop and restart endpoints through systemd, launchd or docker, locally or over SSH. Every action is logged with whether the effect was actually observed.

Measure

Benchmarks with anchored interpretation, and recipes that record how a model is served — version-controlled, so you can answer why a config changed and what it cost.

How it behaves

Honest by construction.

Nothing is contacted until you ask. It shows how many addresses a scan will touch before touching any of them, and refuses public address space by default.
It says when it cannot tell. An endpoint it cannot identify is reported as OpenAI-compatible, not guessed at. Naming an engine from a port number is a coincidence, not evidence.
Secrets are referenced, never stored. The plan holds the environment variable's name, not its value.
An optional agent, on the same rails. It uses the paths the buttons use — same confirmations, same audit log, no privileged route. It can run on your own local model, or on your ChatGPT or SuperGrok subscription.
One web app, one machine. No Kubernetes, no database, no agent installed on the hosts it watches.