Describe an outcome in one sentence. It routes itself to the department that owns it, that head turns it into a plan you approve, specialists build it and a named reviewer signs it off — while you watch every decision happen.
Echlon AI runs on your own machine — this page is just the shop window, so there is no log-in for it to open until it knows the address to send you to.
That does not look like a web address.
On the machine running it, the address is http://localhost:4180. This device
remembers whatever you enter, so Log in goes straight there next time.
A real run, replayed — every message, delegation and review verdict comes from an actual run log.
Coding assistants help one person type faster. They do not plan work, split it, review it, or refuse to ship something broken.
Roles are data, not code: a YAML file and a job charter. Adding one puts it in the org chart, the delegation pool and the 3D office without touching the engine — and adding one with a team under it creates a department that requests can route to. The entire security organisation was added that way. So was every physical-operations department.
You do not pick a team or a mode. A dispatch step reads the goal and hands it to the head who owns that work — and decides whether the run should produce a product or findings.
Your goal becomes a task breakdown with an owner and 2–5 acceptance criteria per task. Nothing is built until you say go — or let it run autonomously and have the department head sign it off.
Every role names the reviewer who signs its work off. Results are checked against their acceptance criteria and sent back when they fall short — before you ever see them.
Lint and tests run as real commands, placeholder scripts that only print a message are rejected, and a UI app must boot and serve. "It compiles" is not evidence.
Tasks are tracked from the plan to a recorded verdict. Stop early and the work goes back in for what remains; a plan still unfinished fails the run rather than being delivered half-done.
Per-phase and per-run spend caps, live cost by model, and one button that stops the whole company — subagents included — mid-task.
The whole chart, hired. Base salaries only — before payroll tax, benefits, equipment or the months it takes to assemble them.
A head staffs only the roles a goal actually needs, so most runs use a handful of the fifty-two — you are not paying for the chart, you are paying for the run. The engine figure is metered; the human one is an estimate the CTO makes per task while planning, which the Dev Lead flags when the delivered work does not match it. Every run shows its own comparison, with the arithmetic on screen and the salaries yours to set.