DEV Community: Jan Luca Sandmann

Claude Code to AI OS Blueprint: Skills, Hooks, Agents & MCP Setup in 2026

Jan Luca Sandmann — Thu, 05 Mar 2026 08:59:46 +0000

I used to treat Claude Code like a fancy terminal buddy. Then I spent a month turning it into a full-blown AI operating system with persistent workspaces, auto-skills that fire themselves, hooks that notify me on Telegram, sub-agents that hand off tasks, and MCP servers that plug into anything.

Suddenly my Claude wasn’t just coding. It was running my entire workflow while I slept.

This is the exact blueprint I wish existed when Claude Code dropped. It’s 100% free (if you already pay for Claude Pro/Max), works locally or in the cloud, and beats the pants off half-baked “agent” toys like Clawbot or Manus for most developers.

TL;DR: You can turn Claude Code into a real aiOS in under 30 minutes with four layers (CLAUDE.md → Skills → Hooks → Agents + MCP). I’ll show every config file, folder structure, and real example. At the end I’ll share the zero-setup cloud version that adds native iOS & macOS apps and 24/7 execution. Because sometimes you just want the OS without managing the hardware.

The Foundation: Turn Claude Code into a Persistent “Computer” First, stop thinking of Claude Code as a CLI tool. Think of it as a virtual machine that happens to speak natural language.

Core setup (5 minutes):

Install the latest Claude Code CLI or VS Code extension (Anthropic’s official one).
Create a dedicated workspace folder: ~/aiOS
Inside it, run claude code init (or just start a new session with, workspace ~/aiOS).
Now add the brain: create CLAUDE.md at the root.

aiOS Core Instructions

You are the operating system for my AI agents.

Every file you create goes in /workspaces/
Use persistent storage. Never forget previous sessions.
Default to deep research + code execution before answering.
When in doubt, create a sub-agent for parallel work. Claude now treats this folder like its own hard drive. That’s step one of turning a chatbot into an OS.

Skills: Give Your Agent Superpowers (No Code Required) Skills are the killer feature nobody talks about enough in 2026.

Create the folder: ~/.claude/skills/

Drop in plain .md files. Claude auto-detects and invokes them when relevant.

Example skill: deep-research.md

`# Skill: Deep Research
Trigger words: research, investigate, latest, 2026
When triggered:

Run 3+ web searches with different phrasings.
Cross-reference with code execution if numbers involved.
Output a markdown report with sources in /reports/
Save raw data in /cache/ Example skill: git-deploy.md`

`# Skill: Git Deploy
Trigger: deploy, ship, push to prod

Always run tests first
Create PR with conventional commits
Notify me via hook when done I now have 12 skills. They turn Claude from “helpful” to “I didn’t even have to ask.”`

Hooks: The Nervous System (Run Code on Every Action) Hooks are shell scripts that fire automatically.

Create: ~/.claude/hooks/

before-tool.hook (runs before every file edit or terminal command)

echo "🛠️ Agent about to touch $(basename "$1")" | terminal-notifier -title "aiOS" -sound default after-success.hook

if [[ "$CLAUDE_EXIT_CODE" == "0" ]]; then echo "✅ Task complete — sending to iMessage" echo "$CLAUDE_SUMMARY" | imessage "me@icloud.com" fi

Make them executable: chmod +x ~/.claude/hooks/*

Now your aiOS feels alive. It pings you, logs everything, even plays a chime when a sub-agent finishes (yes, I did that).

Agents & Sub-Agents: True Orchestration Claude Code 2026 supports spawning sub-agents natively.

Join The Writer's Circle event
In any session just say:

“Spawn a research agent and a coding agent. Research agent: find latest MCP servers for Postgres. Coding agent: build the integration using the research output.”

Or use the new agent flag for permanent workers:

claude code --agent "daily-report" --schedule "0 8 * * *" --workspace ~/aiOS/report

I run three permanent agents:

Research Bot (pulls news every morning)
Code Reviewer (watches my GitHub)
Personal Assistant (handles emails via MCP) This is where it starts feeling like a real operating system.

The MCP Blueprint: Connect Anything (The Real Game-Changer) MCP = Model Context Protocol. Anthropic open-sourced it in late 2024 and it’s now the USB-C of AI agents.

Claude Code understands MCP servers out of the box.

Quick MCP setup (2026):

Clone the official servers repo: git clone https://github.com/modelcontextprotocol/servers.git ~/.mcp/servers
Edit your Claude config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS): { "mcpServers": { "postgres": { "command": "node", "args": ["~/.mcp/servers/postgres/server.js"], "env": { "DB_URL": "..." } }, "github": { ... }, "notion": { ... } } }
Restart Claude Code.

Now your agents can read/write your actual databases, GitHub repos, Figma files, Slack channels — without custom code. One MCP server = instant tool for every Claude-powered agent.

I have 8 MCP servers running. My aiOS literally controls my entire stack.

Real-World Results After 30 Days

Built and deployed three side projects completely hands-off
Saved ~18 hours/week on research & boilerplate
Agents now handle my entire content pipeline (research → draft → SEO → schedule)
Zero context loss between sessions
This isn’t theory. This is my daily driver.

The Zero-Setup aiOS Upgrade (If You Want It Running on web, iOS & macOS Today)
Press enter or click to view image in full size

Look, the blueprint above is powerful and free. But managing local containers, keeping MCP servers alive, and syncing across devices gets old fast.

That’s why I now run the production version of this exact system on computer-agents.com, the real aiOS.

It’s literally the same Claude Code engine, but:

Every agent gets its own isolated cloud computer (persistent files + memory)
Native web app (designed like a literal OS with 12 apps) + iOS app + native Mac app (check agents from your phone at 2 a.m.)
Built-in skills system, hooks, and full MCP Tool Server support
24/7 scheduled + webhook-triggered execution (they actually work while you sleep)
Python + TypeScript SDKs if you want to embed agents in your apps
Start free with 150 compute tokens (enough to test the entire blueprint). No servers to manage. Same Claude models you already love.

Check it out here: https://computer-agents.com

(There even is a direct comparison page vs Claude Cowork. Spoiler: the cloud + mobile version wins for anyone doing real work.)

Final Thought

Claude Code was never meant to stay in the terminal. With this skills + hooks + agents + MCP blueprint, it becomes the operating system for the agentic era.

Build it yourself for full control. Or let computer-agents.com run the production-grade version while you focus on what actually matters.

Either way, your AI coworkers are about to get a massive promotion.

What’s the first skill you’re adding? Drop it in the comments. I read every one.

(And if you’re tired of local setup, just go to computer-agents.com and spin up your first cloud agent in 60 seconds.)

See you in the aiOS. 🚀

Manus AI (Now Meta) vs Computer Agents in 2026: Why Developers Are Flocking to the Independent Platform

Jan Luca Sandmann — Mon, 02 Mar 2026 13:13:33 +0000

Published March 2, 2026

When Manus AI launched in March 2025, it felt like sci-fi had arrived early. Here was a true general AI agent — not just answering questions, but planning, browsing, coding, creating files, and delivering complete deliverables in a full sandbox computer. It went viral overnight. Then Meta acquired the company for a reported $2–3 billion in December 2025, and suddenly “Manus” became part of the Meta family.

Fast-forward to 2026. Manus is still incredibly capable for one-off tasks. But for developers, startups, and teams building production-grade autonomous workflows, a quieter contender has been stealing the spotlight: Computer Agents .

I’ve spent the last six months running both side-by-side on real projects — long-running research agents, scheduled content pipelines, multi-agent coding teams, and customer-support automations. Here’s the unfiltered 2026 comparison.

What Manus AI Actually Is in 2026

Manus positions itself as “Hands On AI” — an autonomous agent that takes a high-level goal and executes it end-to-end inside its own cloud VM. You give it a prompt like “build a landing page for my SaaS and deploy it,” and it opens a browser, writes code, creates assets, and hands you the result.

Strengths

Extremely intuitive for non-technical users.
Powerful browser operator + wide research capabilities.
Nice Meta ecosystem integrations (WhatsApp, Instagram, Messenger, Slack).
Great for ad-hoc creative or research tasks.

Limitations that hurt in production

Ephemeral workspaces — everything resets after the task finishes.
Single-agent only (no native orchestration of multiple specialized agents).
No cron scheduling or webhook triggers.
Opaque credit system (credits expire monthly, no pre-execution cost estimates, no budget caps).
Owned by Meta → your agent conversations and files flow into the broader Meta ecosystem.
No official TypeScript SDK, limited developer tooling.

Computer Agents: The Developer-First AI Coworker Platform

Computer Agents takes a different philosophy: give every agent its own persistent cloud computer that lives 24/7 and can be scheduled, triggered, and orchestrated like a real teammate.

You deploy agents via web, iOS/Mac app, Telegram, or directly from code (Python + TypeScript open-source SDKs). They keep persistent workspaces with files, conversation history, and context that survive for weeks or months. They can wake up on a cron schedule, react to GitHub webhooks, or run multi-agent workflows.

Key capabilities that actually matter for devs in 2026:

Persistent workspaces — files, codebases, research notes survive between runs.
Multi-agent orchestration — sequential, parallel, conditional, map-reduce. Real agent teams.
True automation — cron schedules, webhook triggers (GitHub, Slack, etc. with HMAC verification), email/Telegram notifications.
Open-source SDKs — official npm and PyPI packages, fully open on GitHub.
Transparent pricing — per-token breakdown, pre-execution estimates, budget caps, credits never expire.
Full data ownership — independent company, encrypted, you can delete everything.
Model choice — pick Claude Opus 4.6, Sonnet 4.5, Haiku 4.5, or mix them.
Rich integrations — GitHub, Google Drive, OneDrive, Notion, Slack, Discord, Email, and custom MCP tool servers.

Head-to-Head Comparison (2026 Edition)

Feature	Manus AI (Meta)	Computer Agents	Winner for Devs
Ownership	Meta	Independent	Computer Agents
Persistent workspaces	Ephemeral VMs	Permanent across sessions	Computer Agents
Multi-agent orchestration	Single agent per task	4 strategies + API	Computer Agents
Scheduling & triggers	None	Cron + webhooks + events	Computer Agents
Open-source SDKs	No TS SDK	Official TS + Python (GitHub)	Computer Agents
Pricing transparency	Blind credits (expire monthly)	Per-token + estimates + budget caps	Computer Agents
Data ownership & privacy	Flows to Meta ecosystem	Yours, delete anytime	Computer Agents
Model choice	Fixed mix (Claude + Qwen)	Full Claude family, user selectable	Computer Agents
Developer API & streaming	Basic	REST + SSE + programmatic management	Computer Agents
Mobile/desktop apps	Web + some Meta apps	Native iOS + Mac + Telegram bot	Computer Agents

Real-World Developer Scenarios

Scenario 1: Long-running coding/research project

You want an agent that maintains a codebase over 3 weeks, runs daily research, and updates a Notion dashboard.

Manus: You’d have to rebuild context every time.

Computer Agents: One persistent agent with scheduled runs and searchable history. Wins.

Scenario 2: Production automation

GitHub PR → agent runs tests, writes docs, posts to Slack.

Manus: No native webhooks.

Computer Agents: Webhook trigger + multi-agent workflow. Wins.

Scenario 3: Cost control on a team

10 engineers running agents daily.

Manus: Credits disappear at month-end if unused.

Computer Agents: Balance rolls over forever + budget caps. Wins.

Scenario 4: Building agent features into your own product

You want to expose agent capabilities to your customers via API.

Computer Agents SDKs + persistent threads make this trivial.

Pricing Reality Check (March 2026)

Manus

Free tier
Standard: ~$20/mo (4k credits)
Higher tiers up to $200+/mo
Credits expire monthly — classic “use it or lose it.”

Computer Agents

Free: 150 compute tokens (~15–23 tasks)
Pro: $19–29/mo for 1k tokens
Scale/Max plans with rollover + budget protection

Most teams I know end up cheaper on Computer Agents once they factor in unused credits and transparency.

The Independence Factor

Manus proved the market exists. Meta buying it validated that autonomous agents are the future. But many developers (myself included) are uncomfortable having their agents’ memory, code, and research living inside the Meta advertising machine.

Computer Agents is the independent, dev-owned alternative that launched right as the acquisition news hit — and the timing was perfect. Their /compare/manus page is refreshingly honest and became required reading in several Slack communities I’m in.

Final Verdict in 2026

If you’re a casual user who wants to occasionally generate slides or research a topic from your phone via WhatsApp — Manus is still fantastic.

But if you’re a developer, founder, or team shipping real automation, building internal tools, or integrating agents into products… Computer Agents is the clear winner.

Persistent memory, true scheduling, multi-agent teams, open SDKs, and pricing you can actually predict — these aren’t nice-to-haves anymore. They’re table stakes for production use.

Ready to try the platform that treats agents like real coworkers instead of one-shot VMs?

→ Start free at computer-agents.com (no credit card required)

→ Check their Manus comparison page

→ Import your existing agents — the SDK makes migration surprisingly painless.

The agent era is here. Just make sure you pick the platform built for the long game.

What are you using in 2026? Drop your experiences with Manus or Computer Agents in the comments — I read every one.

Disclosure: I’m a paying user of Computer Agents and have no affiliation with either company. This is based on real usage.

AI Agents That Actually Ship Work: A Practical Guide (With Real Alternatives)

Jan Luca Sandmann — Sun, 01 Mar 2026 07:52:47 +0000

Most “AI agent” demos look great in a 2-minute video and fall apart in production.

The hard part is not generating text.

The hard part is reliable execution across tools, files, retries, and long-running tasks.

If you're evaluating options right now, this guide is the fastest way to avoid expensive dead ends.

What makes an AI agent useful in real life?

A production-ready agent stack needs five things:

Persistent workspaces (state survives after each run)
Tool use with guardrails (browser, code, files, APIs)
Asynchronous execution (works while you're offline)
Observable logs (you can inspect every step)
Clear deletion and data controls (for trust and compliance)

Without these, you don’t have an AI coworker. You have a fragile chatbot wrapper.

The current landscape (and tradeoffs)

There are three common paths today:

DIY/open-source orchestration: maximum flexibility, maximum operational burden
Single-vendor assistant products: easy start, often limited control and extensibility
Managed agent platforms: faster time-to-value if they get reliability + control right

If you’re comparing options, these two breakdowns are useful:

Why many agent rollouts fail

From talking to teams, the failure modes repeat:

Cold starts and inconsistent runtime behavior
Tool calls that look successful but produce no usable output
No reliable handoff between threads/sessions
Weak auditability (“what exactly happened?”)
Unclear ownership/deletion semantics for user data

The result: teams lose confidence, and adoption stalls.

A practical way to evaluate any agent platform

Before committing, run this checklist:

Can it complete a 30+ minute workflow without manual babysitting?
Can it recover from transient errors and continue?
Can users inspect logs and generated artifacts?
Can users delete data and accounts cleanly?
Can you run it across web + mobile + desktop without changing the core workflow?

If the answer is “no” on 2+ items, it will become a support burden.

Where Computer Agents fits

Computer Agents is built around persistent cloud workspaces and long-running execution, so agents can keep working after you close your laptop.

Useful deep dives:

Final take

Don’t choose an agent product based on the prettiest demo.

Choose based on execution reliability, observability, and control.

If you want to test this with real workflows (not toy prompts), start here:

Try Computer Agents