Skip to main content
This guide takes you from zero to a shipped, human-accepted change in under 15 minutes. You will install CyberOS into your repository, write a task that describes what to build, trigger your AI agent, and hold the two human gates that govern every piece of work. No prior CyberOS knowledge required.
1

Prerequisites

Before you install, make sure you have the following available:
  • Python 3.11+ — the CyberOS memory module and doctor command run on Python.
  • Node 24+ — required for the npx cyberos CLI and the MCP server.
  • An AI agent — CyberOS works with Claude, Codex, Cursor, Gemini, Grok, Command Code, Windsurf, Copilot, and any other file-and-shell agent. The Claude plugin gives you the most seamless experience with one-command /ship-tasks access, but it is not required.
Your existing codebase does not need any changes. CyberOS installs entirely under a gitignored .cyberos/ folder and a tracked docs/tasks/ folder — it never touches your source files.
2

Install via npx (recommended)

Run the following command from your repository root:
This single command does everything you need to get started:
  • Vendors the .cyberos/ folder — the CUO workflow engine, the BRAIN memory protocol, the plugin, and the gate runner land under .cyberos/, which is immediately added to your .gitignore.
  • Scaffolds docs/tasks/BACKLOG.md — your task index is created if it does not already exist. Existing files are never overwritten.
  • Creates the local BRAIN store at .cyberos/memory/store/ — your append-only, audit-chained memory store is ready from the first run.
  • Writes agent entry filesAGENTS.md (the canonical cross-tool spine), CLAUDE.md, GEMINI.md, .cursorrules, and per-agent skill directories are created as needed, so every skill-aware agent can drive the workflow immediately.
  • Auto-detects your gate commands — CyberOS inspects your repo for Cargo.toml, package.json, pyproject.toml, go.mod, and other stack markers, then wires build, lint, test, and coverage commands into .cyberos/gates.env automatically.
Install the Claude plugin for one-command /ship-tasks access. In Claude, go to Settings → Plugins → Add and pick the file dist/cyberos/cyberos.plugin. Once installed, typing /ship-tasks in your repo session drives the next eligible task end to end without any additional setup.
3

Or install via terminal (alternative)

If you prefer to install directly from a CyberOS payload you have built or downloaded, run the shell installer instead:
The result is identical to the npx route. Use this option when you are working offline, in a CI environment, or when you want to pin to a specific payload version. You can also run install.sh again at any time to update an already-initialised repo — it backs up gates.env and never overwrites your BACKLOG.md or existing tasks.
4

Verify installation

Confirm that the memory store is healthy and the installation is complete:
A successful install prints a single line:
If you see any warnings or errors, re-run npx cyberos install from your repo root to repair the installation. The doctor command walks the full invariant set defined in memory.invariants.yaml and reports the store’s health in detail.
The .cyberos/ folder is gitignored — never commit it. It is a vendored machine that regenerates on every install run. Your BRAIN store (.cyberos/memory/store/) contains local tenant data that must stay out of version control.
5

Write your first task

Create one file under docs/tasks/. Use the naming convention docs/tasks/<module>/task-<MODULE>-<NNN>-<slug>.md, or copy the template from .cyberos/cuo/templates/. Here is the minimal skeleton:
Key frontmatter fields:
  • id — unique task identifier, used everywhere as the canonical reference.
  • classproduct for new capabilities; improvement for hardening, refactors, and fixes.
  • status: ready_to_implement — marks the task as eligible for the agent to pick up. Tasks in draft are skipped by the queue.
  • priorityMUST, SHOULD, or COULD following MoSCoW convention.
After saving the file, add one line to docs/tasks/BACKLOG.md in the appropriate module section. The task file’s status: field is the record of truth; BACKLOG.md is the index.
6

Trigger the agent

With the Claude plugin installed, type the following slash command in your repo session:
The Claude plugin also supports these commands:Without the plugin, paste this prompt into any file-and-shell agent (Claude, Codex, Cursor, Gemini, Grok, or any other):
The agent picks the first ready_to_implement task from your backlog, implements every normative clause, runs your repo’s machine gates (bash .cyberos/cuo/gates/run-gates.sh), reviews its own work, and advances the task status as it goes. It runs continuously between the two human gates and self-resolves everything it can verify by machine.
7

Hold the two human gates

The agent stops and waits for you at exactly two points. These are the only moments the process requires your attention, and it cannot proceed without you:Gate 1 — Review acceptance (reviewing → ready_to_test)The agent presents what it built together with its code-review findings and an edge-case matrix. Read the task’s normative clauses against the change. If everything is correct, say “approved” and the agent moves the task to ready_to_test. If something is wrong, describe the issue in one sentence — the task routes back to ready_to_implement with routed_back_count incremented, and the agent tries again.Gate 2 — Final acceptance (testing → done)The agent presents its test evidence: gate output, coverage results, and proof that each acceptance criterion passed. If you are satisfied, say “done”. The task status becomes done and the work is complete. If something is still wrong, route it back the same way.The agent never sets done itself. An agent that marks its own work done without a recorded human verdict is broken — report it. Every verdict you record is written as an audit row in the BRAIN with your identity, the task ID, the prior status, and the timestamp.

What’s next?

You have installed CyberOS, written a task, and shipped your first change. From here you can:
  • Read Core Concepts to understand the BRAIN memory model, the full task lifecycle, and the Skills system in depth.
  • Explore the CUO module to see the complete workflow doctrine, execution discipline, and the 10-state status contract.
  • Browse the Skills Library to discover the 104 built-in Skills and learn how to invoke them from any agent.
  • Check the Install Guide for all available installation channels: GitHub Action, Docker, MCP server, Makefile target, and more.