Key concepts
Instance
Section titled “Instance”Your instance is your private ExoHuman deployment: a web app, an API, a scheduler, an MCP server, a PostgreSQL database, a Redis cache and one container per agent. In ExoHuman Cloud it lives at https://<subdomain>.exohuman.cc. An instance belongs to exactly one customer and runs one pinned ExoHuman release. See Instances & architecture.
An agent is a named, long-lived worker. Each agent has:
- a container with a persistent workspace (
/home/developer), its own tools and resource limits; - instructions in
AGENTS.md, plus skills it can use; - a runtime — the agent harness that does the thinking (Pi, Claude Code, Codex, Gemini CLI or OpenCode);
- credentials injected as environment variables and files;
- schedules, loops, canvases, reports and a chat history;
- an owner, who can share it with other people.
Agents start when they have work and are stopped or suspended when idle. See Creating agents.
Run and execution
Section titled “Run and execution”Every unit of work — a chat message, a scheduled task, a delegated request, a webhook — is an execution. Executions have a status, a duration, a cost and logs. Concurrent runs is the number of executions working at the same moment; plans and per-agent parallel capacity limit it.
Compute profile
Section titled “Compute profile”A compute profile describes the resources an agent’s container gets when it runs: vCPU, memory and optionally a GPU. In ExoHuman Cloud you choose a profile per agent and pay per minute of running time. See Compute profiles & GPU agents.
Workspace and Inbox
Section titled “Workspace and Inbox”The Workspace is the home screen: a conversation with your principal assistant plus a rail showing canvases, files, projects, asks and work. The Inbox collects approvals, questions and alerts from all agents you can see. See Workspace and Inbox & approvals.
Principal assistant
Section titled “Principal assistant”Your assistant is a personal agent pinned at the top of the Workspace. It talks to you, delegates to the agents it is allowed to reach and reports back. See Principal assistant.
People, roles and teams
Section titled “People, roles and teams”- Roles decide what a person can do across the instance:
user,operator,creatorandadmin. See Users, roles & teams. - Teams group people. Team leads can be delegated member management and sharing.
- Ownership and sharing: agents are owned by a person and shared explicitly with people or teams.
Permissions between agents
Section titled “Permissions between agents”Agents only talk to agents they have been explicitly allowed to reach. Permissions are directional: allowing A → B does not allow B → A. See Permissions.
Guardrails
Section titled “Guardrails”Guardrails are deterministic safety checks around agent tool use: a deny-list of shell commands, protected files, secret-leak scanning and turn limits. They fail closed. See Guardrails.
Models, gateway and budgets
Section titled “Models, gateway and budgets”Agents call large language models through provider connections. Admins define which models a team may use and a spending budget; the optional managed model gateway routes and meters calls. See Model gateway & budgets.
Release
Section titled “Release”ExoHuman ships as versioned releases (semantic versioning). Your instance is pinned to one release and updated by us in a planned window. See Release notes.