Cabinet AI: We Set Up an Agent Team

AI Summary

What it is: Cabinet is a free, open-source app that turns your existing Claude Code or Codex subscription into a team of named AI agents with their own schedules and memory, stored as plain markdown files you own.

Who it’s for: Anyone already using Claude Code or Codex who wants recurring work done on a schedule instead of re-explaining context every session.

The key idea: We built a real agent in it, named Scout, to run backlink outreach for this site. This is what that setup actually looked like.

Skip if: You want a beginner-friendly chatbot. Start with AI agents versus chatbots instead.

Cabinet is a free, open-source tool for running a team of AI agents instead of one chat window. We set one up for this site’s backlink outreach and want to walk through what it actually does, what it costs, and whether it is worth the setup.

Table of Contents

What is Cabinet?

Cabinet calls itself an “AI-first startup OS.” In practice, it is a local app that sits on top of an AI coding tool you already have, either Claude Code or Codex CLI, and turns it into something closer to a small company. You create named agents, each with a role, a memory, and a schedule. They read and write to a shared set of files on your own computer instead of a database somewhere else.

Hila Shmuel, a former engineering manager at Apple, has built it in public on GitHub since April 2026. The project has picked up roughly 2,800 GitHub stars in that time, which is fast for something this new.

Everything Cabinet touches lives as plain markdown files in a folder, with git tracking every change. Nothing to export, no database underneath. If you stop using Cabinet tomorrow, your files are still just files.

How does Cabinet actually work?

Setup is one terminal command:

npx create-cabinet@latest

That opens a local web app and an onboarding wizard that builds a starter team in five questions: what the cabinet is for, what you want it to accomplish, who the first agent is, what that agent’s job is, and how often it should run. Cabinet ships 20 pre-built agent templates across departments like Marketing, Engineering, Sales, and Operations, so you can start from a template instead of a blank page.

Each agent gets three things: a role written in plain English, a set of markdown files it reads and updates, and a “heartbeat,” which is a schedule (daily at 9 AM, every six hours, and so on) that fires the agent automatically, even when you are not at your computer.

Cabinet does not run its own model. It wraps Claude Code or Codex CLI, so any usage runs through the subscription you already pay for. There is no separate AI bill and no markup.

We built a real agent: meet Scout

Here is the actual cabinet we set up for this site on 4 August 2026.

We named the cabinet “Backlink Outreach & Domain Authority” and gave it one goal: build and run a continuous, white-hat outreach system that earns relevant backlinks through personalized pitches offering real mutual value, link swaps, guest posts, resource-page placements, or newsletter features.

Inside it, we hired one agent named Scout with this role:

Research new link opportunities daily. Maintain living markdown files: prospects.md, outreach-log.md, wins.md, and swap-partners.md. Draft personalized outreach emails and follow-up sequences that offer real value. Always queue drafts for approval before sending. Track replies, placements, and outcomes. Work white-hat only. When in doubt, ask before taking external action.

Scout’s heartbeat is set to run daily at 9 AM. Its first filed task was to create those four tracking files, research an initial list of 10 to 12 backlink opportunities, cross-check each prospect against our Gmail so it never pitches someone we have already contacted, and note a suggested outreach angle for each clean prospect.

Two details stood out in that setup. First, everything Scout drafts sits in a queue for a human to approve before anything is sent. We did not have to configure that as a Cabinet default; Scout followed it because we wrote it into the role, matching the safeguard we already run manually for outreach elsewhere on this site. Second, the four markdown files Scout maintains are readable in any text editor, which means the “state” of an entire outreach campaign is a handful of files you could hand off to someone new with no explanation.

Scout has been running for about a week as of this writing, with no placement count yet worth reporting. We will follow up once there is real data to share.

What does Cabinet cost?

Cabinet itself is free, open source under the MIT license, and self-hosted. There is no subscription and no paywall on the app.

The real cost is whatever you already pay for Claude Code or Codex CLI, since Cabinet runs on top of one of those. A hosted version called CabinetCloud is planned for teams that do not want to self-host, but it remains waitlist-only with no public pricing yet.

What can Cabinet do besides run agents?

Missions and tasks. A Kanban board for breaking a goal into steps and assigning them to an agent.
Internal chat. Team channels where you @mention an agent to trigger it directly.
Embedded HTML apps. Drop an index.html file into any folder and Cabinet renders it as a live dashboard inside the workspace, which is not something a plain note-taking app does.
Git-backed history. Every save auto-commits, so you can diff or restore any page to any earlier point.
Web terminal. A full Claude Code terminal inside the browser, for direct sessions and debugging.

How is Cabinet different from Notion or Obsidian?

Obsidian and Cabinet both keep your notes as local markdown files you fully own. The difference is what happens next. Obsidian does not run agents or scheduled jobs on its own. Notion has AI features, but it is a hosted database, not a folder of files, and it does not orchestrate a team of agents that act on a cron schedule either.

Cabinet’s closest relative is really tools like Zapier or n8n-style automation, except the “actions” are AI agents with judgment and memory instead of fixed if-this-then-that steps.

Who should actually use Cabinet?

Cabinet fits people who already use Claude Code or Codex and are comfortable running one command in a terminal. Setup assumes that baseline; it is not a point-and-click product for someone who has never opened a terminal.

Recurring, well-defined work is the strong-fit case: research sweeps, outreach, weekly reports, inbox triage, competitor monitoring, anything that benefits from the same agent doing the same job every day and remembering what it did yesterday.

Anyone who wants zero maintenance will find it a weak fit. Cabinet is software you run and check in on, rather than a hosted service that runs quietly forever, and it is only four months old: 63 open GitHub issues at the time of writing.

What AI can’t do here

It cannot decide who to contact. Scout can draft fifty outreach emails, but it cannot know which relationship is worth protecting more than a backlink is worth earning.
It cannot own the send. Every draft sits in a queue because a wrong email under our name is our problem, not the agent’s.
Whether the whole campaign is still the right idea stays a human call. Someone still has to look at wins.md every so often and ask if this is working.

New AI Tools, Tested First

Human curated and edited, for beginners who want a finger on the pulse without the hype.

Free forever. Unsubscribe anytime.

Common questions about Cabinet

Is Cabinet free?
Yes. Cabinet itself is free and open source under the MIT license. You pay only for the Claude Code or Codex subscription it runs on top of.

Do I need to know how to code?
Not to configure an agent, but you do need to be comfortable running one command in a terminal to install it, and Cabinet assumes you already have Claude Code or Codex CLI set up.

Where does my data go?
Nowhere, by default. Cabinet stores everything as markdown files on your own machine with git history. The only data that leaves is the model call to Claude or Codex when an agent actually runs.

Can Cabinet send emails on its own?
Only if you configure it to. In our setup, Scout drafts outreach and queues it for approval; nothing goes out without a human clicking send.

Is there a hosted version?
CabinetCloud is planned for people who do not want to self-host, but it is waitlist-only with no public pricing as of this writing.

How is this different from just using Claude Code directly?
Claude Code sessions start fresh and lose context when they end. Cabinet adds persistent memory, scheduled runs, and multiple named agents working from the same shared files.

Sources

Cabinet official site
Cabinet on GitHub (source, README, license)
Our own Cabinet setup record for beginnersinai.org, created 4 August 2026.

Google Wants a Dead Airline’s Data

How to Turn Off AI Everywhere

ChatGPT for Teens: Parent’s Guide