An Agentic Operating System for Life

Your Life OS

What you need across life and work: how the agentic OS covers it

Your marketing: audience research, paid ads, social engagement, brand-voice content

→

Research agents analyze and score audiences. Ad agents build campaigns. Content agents draft in your voice and publish on schedule, all 24/7

Persistent memory: every research finding, every result, every experiment: retained and built on

→

The memory layer that powers your marketing agents. Each run starts smarter than the last. Compounds over months. Extends to every domain you add

Business operations: CRM, lead capture, nurture, pipeline, follow-up

→

The agentic OS can replace your CRM entirely or plug into and extend the one you have. Agents own the workflows

Personal life: scheduling, research, thought capture, recurring personal systems

→

Add agents for personal domains as you need them. Same OS, same Telegram control surface, same memory layer

Desktop and app automation: tools with no API, GUI-only workflows

→

Computer-control agents navigate native apps, fill forms, click, screenshot: anywhere a human could go, the agent can go

Three things that compound over time: 24/7 always-on: your OS never sleeps or loses momentum. Persistent memory: every insight, every experiment result is retained and built on; the next task starts smarter than the last. Self-evolution: agents audit and improve their own work; the fleet gets better without you lifting a finger.

Open vs. locked

Multi-model: no vertical lock-in

The best agentic OS runs multiple LLMs. You are never locked to one vendor's stack.

Mix models per job

Orchestrator: best Opus at highest effort (recommended default). Workers: Sonnet for execution. Bulk: Ollama locally for free. Right model, right task.

Adversarial review built in

One model drafts. A panel of others critiques. You get the refined result: higher quality, caught by multiple perspectives.

Multi-domain fleet

One cortextOS instance runs separate agent teams for different parts of your life, with fully isolated memory and voice per domain.

Open source = portable

MIT license. Your agents, your data, your infra. Swap models as better ones launch. No renegotiation, no migration tax.

Your Life OS (open)

Any model. Any domain. Your data. Runs on your machine. Free forever.

Vertical lock-in

One provider's stack. Their model roadmap. Their pricing. Their migration path.

Model routing reference

Orchestrator Best Opus model at highest effort. This is the recommended default. Runs your chief-of-staff agent: planning, coordination, oversight.

Workers Sonnet for execution tasks. Haiku for read-only lookups and formatting. Right tier, right cost.

Bulk / local Ollama runs on your machine for free. Best for high-volume embedding, classification, and repetitive summarization.

Fable 5 Optional toggle, not the default. Pull it in as the orchestrator only for really long or big-endeavor projects, or when you are willing to spend more on tokens.

🛠

Where the ecosystem lives

James Goldbach's canonical repos

The foundation repos and course path. Zach's fork builds on all of these.

⌄

01

cortextos

The canonical base OS. Actively maintained. Your fork builds on top of this foundation.

github.com/grandamenium/cortextos
02

multimodal-rag

Local memory knowledge base. Ingests video, images, audio, documents. The long-term memory layer, already built and wired in.
03

dream-skill

Memory consolidation. The agent that reflects on past sessions, surfaces patterns, and evolves over time.
04

Agent Architects course (lessons 1-4)

James's build path: Telegram bot then agent soul then keep-it-alive then instant comms + multimodal memory. Four lessons and you have a working persistent agent.

skool.com/agent-architects
+

Also in the ecosystem

m2c1 (autonomous dev loop), autoresearch-anything, social-media-insights, skill-optimizer.

MIT License: what it means for you

Under the MIT License you can:

Use cortextOS in whole or in part, in any product
Modify it freely: change anything you want
Keep your own additions private and proprietary (no copyleft, no obligation to open-source your derivative work)
Build it into a commercial product and sell it

The only requirement: retain the original copyright notice and MIT license text in your copies. One file, trivial. No royalties, no permission needed.

Why this fork

What You're Building and Why to Start Here

The open-source project is called cortextOS, an AI agent framework. But the version you're looking at is a fork that has been built heavily onto over many months: better onboarding, computer-control integration, community skills, persistent memory patterns, and real production patterns from running a live fleet 24/7. Zach recommends starting here rather than the upstream: it skips months of trial and error you'd otherwise have to do yourself.

What makes this different from just using Claude in a browser: your agents are persistent. They have memory, identities, scheduled tasks, and Telegram as a control surface. You text your agent and it acts. It remembers what happened last session. It runs while you sleep, and reports back when it finishes.

The trajectory: You start with your marketing. A few months in, you have an agent for content, one for research, one for scheduling, one that monitors your business metrics. Then it grows beyond work. That is how Zach uses his: it became the OS for all of life. That is what cortextOS scales into.

Set expectations early

What It Costs to Run

This is a subscription-based power tool. The costs below reflect running two businesses plus a personal Life OS: constant content, graphic generation, and a full agent team running around the clock. Your starting setup will cost less. But you should know the ceiling before you begin.

Subscribe via your computer's web browser, not the iOS App Store. Apple takes ~30% of every subscription through the App Store. The same plans cost significantly more on iPhone. Go directly to each provider's website to subscribe.

Service	Plan	Monthly	Notes
Claude (Anthropic)	Max	~$200	Zach's recommendation for heavy users. Powers all your agents. Subscribe at claude.ai, not iOS.
Codex (OpenAI)	Pro	~$100	Powers computer-control and desktop automation. Subscribe at openai.com, not iOS.
Gemini (Google)	Advanced / API	~$50	Knowledge base embeddings and research queries. Free tier works to start; upgrade as you scale.
Required core		~$350/mo	Claude Max + Codex Pro + Gemini. Everything you need to run the OS.
Agent Architects Skool (James Goldbach)	Optional	~$100	Highly recommended. The community your OS learns from: Forge scrapes James's latest methods and content to keep your system sharp. Subscribe at skool.com.
Full recommended setup		~$450/mo	Core + Skool community. Zach's personal power-user total.

You don't need all three on day one. Start with just Claude Max. It powers everything through the initial setup. Add Codex when you want desktop automation. Add Gemini when you build a knowledge base.

Minimal human steps, in order

Your Setup Walk-Through

Every step below is tagged: Human means you do it directly. Agent-Automated means your agent handles it once it has the access it needs. The human steps are deliberately few.

You do these steps

First, before everything else

Human

Set up 1Password: your credential vault

Every service your agent accesses (Telegram, Google, APIs, cloud providers) needs credentials. Your 1Password vault is where those credentials live. Your agent reads from it at runtime so nothing ever sits in a plaintext file. Set this up first so every secret has a home before your agent starts asking for logins.

1

Install the 1Password desktop app: 1password.com/downloads

Create your account. This is the master vault for your entire OS.
2

Install the 1Password CLI (the op command)

Mac: brew install 1password-cli then op signin. This is how your agents read credentials without a GUI: they call op item get at runtime.
3

Create an "Automation" vault inside 1Password

Agent service credentials go here: Telegram bot tokens, API keys, service accounts. Keep them separate from your personal logins.

Never paste secret keys or passwords into the chat

This is a hard rule for your protection. Never type an API key, token, or password into the Claude Code chat window. Your agent retrieves credentials by calling 1Password directly. If it ever asks you to paste a secret into the chat, that is a mistake. Save the secret to your 1Password Automation vault instead, and tell your agent the item name.

Every time your agent needs to authenticate to a new service, here is how it works:

How auth works: the pattern you will use every time

Your agent: Opens the browser next to you and navigates to the login screen
You: Type your credentials into the real login form: not into the chat
Your agent: Guides you to save the credential directly into your 1Password vault
Your agent: From that point, retrieves it from 1Password on its own: no more manual entry

Second: grant the keys

Human

Computer-control and browser access grants

Your agents can control your Mac: navigate browser tabs, interact with desktop apps, take screenshots, and manage files. That power only works after you grant the right system permissions. If you skip this step, half your agent's workflows will hit dead ends.

A

macOS Accessibility + Screen Recording

System Settings > Privacy & Security > Accessibility: add your terminal and PM2. Do the same under Screen Recording. Grant these to PM2, not just your terminal app. Your agents run under PM2 in production and need the grants to reach the screen.
B

cmux: your terminal + browser panel workspace

This is where you'll work alongside your agents: a terminal pane and a browser panel side-by-side. Your agents open URLs in the cmux browser directly, not in your Mac's default browser. Set up your cmux layout before starting your agents.
C

Chrome DevTools browser control

This lets your agents navigate pages, fill forms, click buttons, take screenshots, and scrape data. Add the Chrome DevTools MCP server to ~/.claude.json when you need it; remove it after each session to keep memory clean.

Do steps 1 and 2 before running the installer. Once your agent has browser access and 1Password wired up, it can drive the credential-saving steps on your behalf through the rest of setup.

Third: install the OS

Human (one command)

Run the one-line installer

Open a terminal. Paste this one line and press Enter. The installer handles everything from here automatically.

# Mac / Linux
curl -fsSL https://raw.githubusercontent.com/drzachconner/cortextos-memily/main/install.mjs | node

# Windows (PowerShell)
node -e "$(irm https://raw.githubusercontent.com/drzachconner/cortextos-memily/main/install.mjs)"

Your installer automatically checks for and sets up:

▸
Node.js 20+, npm, git, Homebrew (Mac), jq
All runtime prerequisites. Xcode CLI Tools installed if missing.
▸
Claude Code: npm install -g @anthropic-ai/claude-code
Prompts you to run claude login if not already done.
▸
PM2 process manager: keeps your agents running 24/7
▸
Clone repo to ~/cortextos, build, link global CLI
▸
Run cortextos install: sets up your state directories and config

Windows: WSL (Windows Subsystem for Linux) is required. If you don't have it: run wsl --install in an admin PowerShell, restart your machine, then re-run the installer.

Fourth: start the wizard

Human (two commands)

Open in Claude Code, run /onboarding

Once the installer finishes, open your cortextOS project in Claude Code and launch the onboarding wizard:

# Opens your project in Claude Code
claude ~/cortextos

# Inside Claude Code: launches the guided wizard
/onboarding

Those are your last two manual steps. The wizard walks through everything from here in plain English: you answer questions, it configures your system.

The installer's final output reminds you of exactly this: open the folder, run /onboarding. You don't need to memorize anything.

Your agent takes over from here

What /onboarding does

Agent-Automated

The wizard: what happens step by step

When you run /onboarding, your agent walks through each of these in order. You answer a few plain-English questions; it writes all the config files itself.

1

Detect your runtime auto

Reads config.json to set the right runtime. claude-code is the default.
2

Check onboarding status auto

Looks for the .onboarded flag. Safe to re-run if you get interrupted.
3

Create your Telegram bot via @BotFather you

Your agent opens Telegram alongside you. You message @BotFather, run /newbot, pick a name and username, and copy the token. Your agent saves it to your 1Password vault.
4

Get your chat ID auto

Your agent messages the new bot and fetches your chat ID via the API automatically.
5

Define your agent's identity, soul, goals, guardrails you answer

The wizard asks: what is your agent's name, what is its role, how autonomous should it be, what are its starting goals? Your answers get written to IDENTITY.md, SOUL.md, GOALS.md, GUARDRAILS.md.
6

Write your .env with bot credentials auto

Your agent writes BOT_TOKEN, CHAT_ID, and ALLOWED_USER: pulled directly from your 1Password vault.
7

Configure your crons auto

Sets up your recurring workflows (heartbeat, morning review, and others) as persistent external crons that survive restarts automatically.
8

Write .onboarded flag + send your first Telegram message auto

Your phone buzzes. That is your agent saying hello for the first time. Every future restart skips onboarding and goes straight to work.

The moment it flips

Agent-Automated

Handoff: you step back, your agent runs

🤝

Your handoff point

The moment .onboarded is written to your system

Everything before this line required your direct input. Everything after, your agent drives on its own.

What you did (steps 1-4):

✓
Set up your 1Password vault, CLI, and Automation vault
✓
Granted Accessibility and Screen Recording to PM2
✓
Set up your cmux workspace and browser panel
✓
Ran the one-line installer
✓
Opened your project in Claude Code, ran /onboarding, answered the wizard's questions

What your agent does from here, without you:

▸
Runs your heartbeat and memory crons on schedule
▸
Accepts and executes your tasks via Telegram
You text it, it acts. That is the whole model.
▸
Writes to memory files after every session
Picks up full context from prior sessions (no cold start).
▸
Auto-restarts via PM2 on crash and resumes from last state
▸
Context-rotates cleanly every ~71 hours with no data loss
▸
Your local dashboard updates live with tasks, agent health, and activity
Run cortextos dashboard to open it. Runs entirely on your machine: your data stays yours.

Building your fleet

Your Agent Roster: What Each One Does

You start with one agent. As your OS grows, you spawn more, each with a focused role. Your agents communicate via a shared message bus and can hand tasks off to each other. Below is the full roster, genericized to roles you can adapt to your life. The names are examples from a running fleet: you can keep them or rename them to anything.

🧭

Sage

Orchestrator / Chief of Staff

The primary agent you talk to. Routes tasks to specialist agents, maintains your high-level context, manages your agenda, and synthesizes results. Your single inbox: everything flows through Sage.

🗺

Atlas

Analyst / Fleet Intelligence

Monitors your fleet health, efficiency, and system performance. Runs audits, surfaces trends, and identifies where your OS can improve. Proactively flags inefficiencies before they cost you.

🛠

Forge

Engineering / Automation

Owns all code and integration work. Builds custom scripts, wires up APIs, handles infrastructure plumbing, and debugs technical issues. Your engineering co-pilot.

✍

Scribe

Content / Brand Voice

Writing, editing, email drafting, social captions, blog posts, and guides. Holds your brand voice and ensures every piece of content sounds like you, not an AI.

🚀

Mercury

Social / Publishing

Social media management, post scheduling, platform publishing, and outreach workflows. Handles the mechanics of getting content from draft to live.

📚

Librarian

Knowledge Base / Research

Ingests documents, transcripts, URLs, and notes into a searchable vector knowledge base. Retrieves relevant context on demand. Your long-term memory for information: it grows smarter the more you feed it.

⚡

Felix

General Worker

A flexible general-purpose worker. Takes overflow tasks from specialist agents, handles one-off jobs, and processes batch work without clogging your specialist queues.

📊

Hippocrates

Data / Stats / Audits

Data analysis, metric tracking, KPI audits, and reporting. Pulls data from connected sources, finds patterns, and flags anomalies. The numbers agent.

🖥

Argus

Desktop / Computer Control

Controls desktop applications directly: clicks, types, navigates native apps, takes screenshots, and drives GUI workflows that have no API. The computer-control specialist.

🔭

Apollo

Domain / Client-Scoped Agent

A focused agent scoped to a specific domain, client, or project. Has its own Telegram bot, its own identity, and its own guardrails. Useful when a context needs a dedicated presence separate from your main OS.

You don't need all ten on day one. Start with Sage. Add Forge when you want engineering automated. Add Scribe when you want a consistent writing partner. Build your roster as your actual needs emerge.

James Goldbach, Agent Architects

The mindset + your first-week priorities

Your OS is built on the Agent Architects methodology. A few principles that will make the difference between a working fleet and a weekend experiment:

Start with one agent and run it for a week. Resist the urge to spawn five agents on day one. Get one agent doing real work (real tasks, real memory, real Telegram back-and-forth) before you add more. Your fleet compounds from there.
Give it a real job, not a demo job. Your agent should handle something you actually care about: your content schedule, your client research, your follow-up system. Toy tasks don't surface the real friction.
Memory is the multiplier. An agent that writes to memory after every session gets smarter over time. An agent that doesn't starts over on every restart. Take the memory protocol in the onboarding seriously: it is non-negotiable.
Telegram is your control surface. Everything you want your agent to do, you send via Telegram message. You don't open the code. You don't run commands. You text it, it acts. That is the whole model.
Ship one complete thing before starting the next. Define done before you start. A feature is not done until it is wired end-to-end: source to storage to visible result. Never have two things half-open at once.
Identity shapes quality. An agent with a clear identity, soul, and guardrails produces better, more consistent work than a blank slate. Take your time in the onboarding wizard: the answers you give shape how your agent thinks and behaves.

James teaches all of this live in the Agent Architects Skool community. It is the right place to see what other people are building, get unstuck, and go deep on the patterns. Joining early is worth it.

Important early move

Join Agent Architects: Learn the Method with James

James Goldbach's Skool community for people building autonomous agent operating systems. The patterns, the troubleshooting, live examples of what others are running: it is all here. Joining before you need it is the right call.

memily claude-code cortextos life-os agent-architects telegram-bots pm2 1password