ApexCorpOperating a company
ApexCorp runs an orchestrator that holds your company and does the work, and a control panel you watch it through. This is what you do once it is running.
Founding a company
The panel asks three things: what the company is called, what it does, and what it is aiming to make each month.
The description is the structural input. Teams, leaders, and a QA team sized to headcount are planned from it, so “ships a developer tool, needs engineering and inbound support” and “runs experiments and publishes findings” produce genuinely different organisations. Write it the way you would describe the business to a new hire.
The monthly objective is a target the company works toward. It is not a budget — spend limits are separate, and deliberately so.
Approving the structure
The Manager drafts an org chart and files a note asking you to approve it. It cannot adopt its own proposal; approval takes a person’s name by design. Until you sign off, the company is running on a shape nobody agreed to, and the deck says so in the top bar.
Open the Teams view, read the roles and reporting lines, and approve or reject with a reason.
Giving it work
The Things to do view takes a title, optional detail, and an urgency flag. Work enters a backlog rather than starting immediately, and a fixed number of agents run at once.
- Ordinary work queues in the order it arrived, and the panel shows each item’s real position — “next but 3” rather than a status word that tells you nothing.
- Urgent work is taken next. It is a position in the queue, not a licence: every limit still applies, and if everything is urgent the order is simply first come, first served.
- A longer brief routes to a stronger model. Detail costs money — usefully, but knowingly.
Reading what it is doing
- Cluster
- The company as a rotating 3D graph. A pulsing bead means that agent is mid-run. Open anyone to see their work, memory and relationships.
- Live feed
- Every dispatch, tool call, memory proposal and escalation as it happens, filterable to memory or escalations alone.
- Org chart
- Declared and observed side by side — what the chart says, and who actually worked with whom. The gap is usually the interesting part.
- Notes
- Everything waiting on a person, in one queue. Resolved by you, never deleted.
Keeping it inside the lines
The limits start tight. Loosen them once you have watched the system run for a while.
- Concurrency ceiling
- How many agents run at once. This is a cost control as much as a scheduling one: without it, ten tasks in ten seconds means ten agents on the model simultaneously.
- Spend ceiling
- Per run and per day, metered against real token pricing. Crossing it halts every agent and files a note rather than continuing quietly.
- Iteration cap
- The maximum reasoning steps in a single run before it stops and escalates.
- Kill switch
- Halts everything immediately. Work you add while halted queues rather than being lost, and resuming starts it.
Notifications
Email and WhatsApp, off until you turn them on, tuned per channel and per event in Settings — work started, work finished, or only when something needs you.
WhatsApp requires a message template approved by Meta before anything can arrive, because business-initiated messages cannot be free text. We handle that during onboarding.
These notify you. The recipient is fixed in configuration and no agent can reach the sender — it is the system telling its operator something, not a way for agents to message anyone.
What it costs to run
Model spend dominates, and the router is what keeps it down. A measured example: a one-line fix, including the QA review it triggers, came to $0.0164. Work classified hard goes to the top tier and costs meaningfully more — under-powering deep work costs more in retries than the upgrade does.
Deployment
Most customers run on our managed infrastructure — you connect a model provider account and we operate the orchestrator.
To run it inside your own network, take the Enterprise tier. Once terms are agreed we send you the full setup instructions by email — requirements, configuration, the safety limits and what to watch on the first few runs — and stay available while you stand it up. It is not a public download, because an autonomous system that runs commands and reads files should be configured by someone who has spoken to us first.
Either way the model credentials are yours, held in the orchestrator’s environment, and never exposed to the browser.