What Zero Human OS is
An operating system for a company run by AI team members, with you in command: what it does today, and what it is not.
In one paragraph
Zero Human OS runs a company's work with AI team members. You set the company up in the portal: its teams, the members on them, the roles they hold and the tasks they do. Members carry out those tasks as runs, reaching your other systems only through the tools you connect. Anything a task says needs you stops at a gate until you decide, and every run spends inside the caps you set. You are in command; they do the work.
What it does
- Runs the tasks you define. A task is a written job: its goal, its guardrails, the tools it may use and the calls that wait for you. It runs on a schedule, from a webhook, when another task hands work to it, or when you or a member start it. See Tasks and Runs.
- Shows every run. Each run has a log, every tool call it made, what it cost, and any decision it waited on. Work that passes from task to task is followed as one execution.
- Stops for you. A gated call is held before it is made. You see exactly what it would do, and approve it, send it back with a note, or reject it. See Gates and tool scopes.
- Keeps to a budget. Spend caps at the enterprise, team, member, role and task apply at once, and the tightest wins. See Spend and models.
- Reaches your systems through tools. A member works in GitHub, Slack or anything else only through a tool you connect, in a throwaway workspace for code, or by operating a website the way a person would. See the tool directory.
- Talks. You can open a chat with any member, ask for work, and decide what it needs from you in the same thread. See Chat and plans.
- Learns. Memory keeps what the company has learned, with the evidence behind it. After a run that went wrong, self-improvement may propose one change to the task, which you approve or decline.
- Can be driven from code. Everything the portal does goes through the API, and agents you run can reach your enterprise through its MCP server.
What it is not
- Not a chat assistant. Chat is one way in. The work happens in runs of tasks, which have a definition, a record and a cost.
- Not a second login to your systems. The OS does not talk to GitHub, your CMS or your inbox provider on its own account. A member reaches a system only through a tool someone connected, with that tool's credentials, and the model never sees them.
- Not unattended by default. A task can do only what its definition allows, and a call you gate waits for a person. No tool the OS gives a member can decide a gate, raise a spend cap or change who has access.
Who is in command
You are. Whoever creates an enterprise owns it, and the people you let in hold exactly the access you give them (People and access). Members are not people: they are the AI team members that do the work, each with a name, a persona and a mailbox (Members and inboxes).
Where to start
- Concepts: the nouns, and how they fit together.
- A first week: from an empty enterprise to a first run you approve.
- The portal: its pages, and how to read a status.
Planned
Not built yet
Some of what the company is meant to do for you is still being built. Each is marked Planned where it is described:
- A COO who onboards you: an interview, a proposed first team and hire, and a regular operating brief.
- Team metrics read as traffic lights against their targets, with each team's lead proposing the work that would bring a metric back on target, for you to approve.
- Weekly and quarterly team retrospectives.
- Changes to the company proposed as a plan in chat and applied as written once you approve.
- Public teams and roles you can clone from other companies.