Ollsoft Company OS

An AI-native company workspace where the Linux kernel is the permission system. Documents and tasks live as plain markdown files on your own server, and the internal tools your team builds live beside them as HTML artifacts. Agents work in all of it as the person who asked, seeing exactly what that person sees.

Your data is never in somebody else's cloud, and an agent reaches it the way a colleague does rather than through an API.

Think Notion, but the permissions go all the way down to the file, the tools are ones your team builds rather than buys, and agents work there beside you. It is an actual operating system, so it is also where the work gets done, not only where it gets written down.

The board on the left is stored in the markdown file on the right. Edit either; both are the same task. A person, a dashboard and an AI agent all work on the same file.

Free for up to three named users, and free for any number of users for sixty days while you evaluate it. Above that, production use in an organisation needs a commercial licence: [email protected]. The source is public either way: read it, audit it, build it, run it for development or testing. Every version becomes Apache 2.0 four years after its release. Details in LICENSE.

Install it with your AI agent

An AI agent on your own computer does the whole thing over SSH: server, hardening, domain, accounts, search, agents, monitoring and backups. It asks one question at a time and checks every step. At the end it moves your existing documents in from Notion, Obsidian, Confluence, Drive, SharePoint or git.

git clone https://github.com/Ollsoft-ai/ollsoft-company-os.git
cd ollsoft-company-os
claude        # then type: /install-company-os

• Claude Code picks the skill up from this repo (.claude/skills/install-company-os/).
• Codex, Hermes Agent or any other agent: start it in the cloned folder and say "Read .claude/skills/install-company-os/SKILL.md and follow it."
• You need ssh (built into macOS, Linux and Windows 10+), about an hour, a card for the server and, optionally, a domain.
• Interrupted? Run it again; it keeps a state file and resumes.

Prefer to run the commands yourself? See Install by hand.

Design

Four primitives. Everything else is what falls out of them.

1. Files. Plain markdown in a git repo at /srv/kb. A document is a file, a project is a folder. Nothing is locked inside a database.
2. The Linux kernel. One OS user per human, and every web request served by a process running as that user (runuser), so the kernel, not application code, decides what each person may read and write.
3. Postgres. A disposable index of those files: drop it and rebuild it from the markdown whenever you like. Row-Level Security mirrors the same Unix permissions, so search and SQL can never return a file you could not cat.
4. HTML artifacts. A company's own small tools (a board, a CRM, an invoice generator) are single HTML files sitting in a folder. They run sandboxed, as the person who opened them, reaching files and SQL through a bridge bound by that person's permissions.

What those four give you without being asked twice:

• The indexer parses the markdown into rows and carries the Unix permissions across with it. A - [ ] task @someone #tag written anywhere becomes a to-do list, and every search is permission-scoped without a line of permission code.
• Sharing is a Unix group. Giving a folder an audience is giving it an owning group, which the file tree, the editor, search, SQL and the agents already obey, because they obey the kernel.
• Multiplayer editing is a CRDT over the same file; the daemon writes it back preserving owner, group and mode, so a shared edit cannot launder permissions.
• Version history is git, attributed to the OS user who made the edit.
• Agents run as the user too, so Claude or Codex see exactly what that person sees. No integration to grant, no second set of credentials to leak.

There is no permission table in this codebase. That is the whole point.

---

What it looks like

Two people, two browsers, one file. On the left somebody types into kanban.md; on the right a colleague's board already carries the change. The board is not synced with the file. It is the file.

The same document open by two people, each cursor named and coloured. Merging is a CRDT over the file on disk, so vim over SSH is a third seat at the table.

A real shell in the browser, as your own Linux account, and grep finds the same card the board was showing. Same files, same permissions, no API in between.

Every open task from every document you are allowed to read, blocked work first. Nobody maintains a second task list.

A CRM is an artifact in a folder. It runs as the person who opened it, so it can only reach what they could.

…and so is an invoice generator, which writes its PDF back beside the contract it came from.

Requirements

Ollsoft Company OS needs a whole machine: a VM or bare metal running Ubuntu 24.04.

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗