A tool that only runs on your laptop is a prototype. A tool that runs without you is an asset — and a liability.
Most CEO Tools Should Stay Local
The premise of this chapter isn't "here's how to deploy your tool to the internet." It's: most CEO tools built with Claude Code should never go to the internet, and don't need to.
The tools described in Chapter 5 — weekly summaries, meeting processors, data cleanup scripts, interview prep briefs — are personal tools. They run on your machine, produce files you read, and serve your workflow. They don't need a server. They don't need a URL. They don't need users.
This is a feature, not a limitation. A tool that stays local is:
- Easier to build
- Easier to debug
- Free to run
- Not exposed to the security risks that come with internet access
Before you think about making something live, ask whether it needs to be. Most of the time, the answer is no.
The Three Levels
| Level | What it means | Risk | Who it's for |
|---|---|---|---|
| Local | Runs on your machine, you trigger it manually | Low | You only |
| Scheduled | Runs on a schedule automatically, outputs go to a file or email | Medium | You, automated |
| Live | Runs on a server, accessible via URL, other people can use it | High | Requires developer review |
Most CEO use cases live at Level 1 or Level 2. Level 3 is where you need a developer.
Level 1: Running It Manually
You built a script. You run it when you want it. The output goes into a folder on your computer.
This is the default. It's where you start. It's sufficient for most personal tools.
The only infrastructure you need: a folder structure that makes sense to you, and a README that explains how to run it. (Have Claude Code write the README.)
Level 2: Scheduled Automation
Your tool runs automatically — every morning, every Monday, every time a file arrives in a folder — without you triggering it manually.
This is appropriate when:
- The output is low-stakes (a draft email, a summary file you'll review)
- You're not writing to any external system automatically
- You've tested it enough to trust it runs correctly without you watching
How to schedule on a Mac or Linux machine:
Claude Code can write a cron job — a scheduled task — for you. Brief:
Create a cron job that runs this script every Monday at 7am
and saves the output to /output/weekly-summary-YYYY-MM-DD.txt
How to schedule on Windows:
Ask Claude Code to set up a Task Scheduler entry. Same result, different mechanism.
Important constraint: Scheduled tools should produce outputs for you to review, not take autonomous action. A script that emails you a draft every Monday is fine. A script that emails your team every Monday without your review is not.
Level 3: Going Live
"Live" means your tool runs on a server and is accessible via a URL. Other people can use it. It runs without you present.
This is where the risk profile changes significantly.
Research from a December 2025 assessment of major AI coding tools found 69 security vulnerabilities across 15 test applications — around half a dozen rated critical. A documented incident: an AI-generated script mistakenly deleted an entire production database. Another: an app deployed with all security logic on the client side, which one security researcher described as "a massive mistake in architecture."
These aren't arguments against building. They're arguments for knowing where the boundary is.
Level 3 requires a developer when:
- Other people will use it (not just you)
- It takes user input (a form, a login, a search)
- It connects to your production systems
- It runs continuously rather than on demand
- It stores data beyond temporary output files
The checklist from Chapter 9 applies here in full.
Anthropic's Own Answer to This Problem
In January 2026, Anthropic launched Claude Cowork — described as "Claude Code for the rest of your work." It connects to Google Drive, Gmail, DocuSign, Slack, and 13 enterprise tools via plugins.
By February 2026, it added enterprise connectors and plugin marketplaces that let organizations build controlled internal tools without managing servers.
If your use case is "share something with my team," Cowork may be the right answer — a managed environment with Anthropic handling the infrastructure and security layer. This is different from deploying a custom tool to the internet yourself.
Keeping It Safe: The Git Habit
Whether your tool stays local or eventually scales, use git from the start. Git gives you:
- Version history — you can see what the code looked like before you changed it
- A recoverable state — if a change breaks something, you can go back
- A record of what was built and when — useful if you ever hand this off to a developer
Claude Code can initialize git for any project and make commits automatically. Add to your CLAUDE.md:
Always initialize git in new projects. Commit after each working version
with a clear message describing what changed and why.
This costs nothing and gives you a recovery option for every version of every tool you build.
When Your Tool Has Outgrown Claude Code
The clearest signal: you started thinking about users other than yourself.
When a tool is just for you, Claude Code's limitations are your problem. When a tool is for other people, its failures are their problem — and potentially your liability.
The handoff signals:
- Someone else needs to use it regularly and depend on it
- You want it to handle customer data, payments, or authentication
- You need it to be reliable when you're not watching
- It needs to scale beyond a handful of uses
At that point, what Claude Code has built is a prototype — a very good prototype, with working logic and a clear interface. Take it to your development team. They'll know exactly what to build because you've already shown them what you mean.
Chapter 9 has the complete signal list.
Hosting Options for Level 3 (When You Work With a Developer)
For reference, the options your developer will likely evaluate:
| Option | Best for | Complexity |
|---|---|---|
| Railway | Simple scripts, APIs | Low |
| Render | Web services, scheduled jobs | Low-medium |
| Fly.io | More control, global deployment | Medium |
| Vercel | Frontend + serverless functions | Low (frontend-focused) |
| Your existing infrastructure | If your company already has servers | Varies |
These are starting points for the developer conversation — not something you should configure yourself without technical review when user data or production systems are involved.
Further Reading
- Fortune: Anthropic launches Cowork — a file-managing AI agent
- VentureBeat: Claude Cowork is coming for the enterprise
- IT Pro: Vibe coding security risks and how to mitigate them
- InfoWorld: Vibe coding tools prone to critical security flaws (Dec 2025 study)
- Aragon Research: Anthropic Democratizes Agentic AI with Claude Cowork
Next: Building the Practice