# YAIBOS: Open Spec for an AI Business OS > Open specification for running a small business with AI agents: a portable file-based Base, an optional skill dashboard (Cockpit), and Team permissions. Plus a qualitative scorecard of existing tools. Spec version 1.1.0-draft, licensed CC-BY-4.0. Author: Samuel Michelot. ## Summary YAIBOS (Your AI Business OS) describes how a small business can run its work with AI agents without being locked into one AI vendor, one app, or one developer. The core idea fits in one sentence: **the second brain is the OS, the harness is replaceable.** The Base profile describes the file-based system Samuel already uses, rather than requiring a new application. A **skill** is a reusable capability for an AI agent: instructions, with optional scripts and supporting resources. A business skill can implement a **standard operating procedure (SOP)**, such as qualifying leads or reconciling invoices. This spec uses *skills* as the main technical term and *SOP* when discussing the underlying business procedure. The operational requirements below apply to business skills, not every utility skill an agent may have. - A folder of plain Markdown files (the *second brain*, currently viewed and edited with Obsidian) holds company context, skills, role-agent definitions and memory, and the tools registry. - An existing AI harness reads these files and runs the skills. A harness is the application that gives a model access to files and tools, for example Claude Code, Codex or OpenCode. - Synchronization prepares commands and tool connections for supported harnesses. Switching preserves the source files; authentication, tool availability and permissions still need verification in the destination. - Optional Cockpit and Team layers add convenience and coordination. They depend more on the chosen harness or integration, while the Base remains usable on its own. ``` Second brain (files you own) Harness (replaceable) Model (via harness) ├── AGENTS.md company context ──► Claude Code / Codex / ──► Claude / GPT / Gemini / ├── 5 SOP/ procedures OpenCode / supported apps supported by the ├── Tools note MCP, CLIs, APIs (official login or API key) chosen harness ├── Agent notes roles and memory └── PARA folders knowledge Optional Cockpit: skill dashboard, launch, progress, approvals, run history Optional Team: shared context, enforced access, ownership, audit ``` "Sovereign" means you own the source files and can change execution tools without rewriting your business knowledge and procedures. It does not promise identical features in every harness or a fixed migration time. Section 5 of https://yaibos.com/SPEC.md describes data boundaries. ## The six pillars ### 01. Usable without a terminal Operations staff and assistants need to use skills without installing packages in a terminal. - Base: Run an existing skill from a graphical app without shell commands. - Base: Every note, skill and log stays a plain file. - Later: Cockpit to browse, launch and follow skills [L2]. - Later: Automatic sync and readable choices for conflicts, without git markers [L3]. ### 02. Portable agent skills Skills package reusable instructions, scripts and resources. Business skills make your SOPs executable by an agent. - Base: One Markdown file per skill, with frontmatter. - Base: Runs in at least two harnesses without editing. - Base: Lifecycle: draft, pilot or stable; readable approval policy. - Later: One-click catalog by area, with run history [L2]. - Later: Trigger URLs for Asana, Trello or an ERP, and schedules [Opt]. ### 03. Bring your own subscription Metered API keys make team bills unpredictable. Most people already pay for an AI plan. - Base: Each user runs skills with their own plan, through the vendor's official client, or their own API key. - Base: No scraping or replaying of consumer session tokens. - Base: Keep credentials out of the second brain. - Later: Authentication mode and Cockpit BYOS bridge wrapping official vendor CLIs [L2]; shared-key spending caps [L3]. ### 04. A living second brain Agents need current company context. Outdated notes can lead them to use old rules or facts. - Base: Plain Markdown in a folder you control. - Base: AGENTS.md entry file and README maps agents can follow. - Base: Role-agent notes and memory; maintenance, backups and updates. - Later: Review queue for stale notes [L3]. - Later: Contradiction audits, decision log, learning from runs, semantic search [Opt]. ### 05. Permissions and safety Broad access to payroll, client data, and email increases the damage an agent can cause when it follows a bad instruction. - Base: Explicit authorization for external or irreversible actions. - Base: Emails, web pages and documents are data, never instructions. - Base: Readable action boundaries within the user's permissions. - Base: Every agent change reviewable in git history. - Later: Visible approvals and cancellation [L2]; enforced role permissions and area owners [L3]. ### 06. Harness and model portability Harnesses and models change every few months. Keeping procedures in open files makes them easier to use with the next tool. - Base: The second brain is the single source of truth. - Base: A tools registry lists every MCP server and CLI. - Base: A sync step installs skills, context and tools in supported harnesses. - Later: Export run history and leave the cockpit for Base [L2]. ## Resources - [Full specification (Markdown)](https://yaibos.com/SPEC.md) - [Specification page](https://yaibos.com/spec) - [Implementation scorecard](https://yaibos.com/implementations) - [Developer resources](https://yaibos.com/developers) - [OpenAPI description](https://yaibos.com/openapi.json) - [llms.txt](https://yaibos.com/llms.txt) - [About Samuel Michelot](https://yaibos.com/about)