# Harbor — A home for the tools your agents build

Like Vercel, but inside your company. Any agent builds it. Your whole team uses it. Admins stay in control.

Harbor is in private preview. Join the waitlist: https://tryharbor.dev/

## The problem

Anyone can build a tool with AI now, but it has nowhere to live and run.

## The solution

Harbor gives AI-built tools a home, with everything they need to run. Every teammate and every agent can use them, securely, in your own cloud.

## What Harbor is

Harbor is the place inside a company where software its AI writes goes live. It is open-source software a company installs once, in its own cloud or on one machine.

Any agent (Claude, ChatGPT, Copilot, Cursor, Codex) can build an app and publish it to Harbor. Harbor gives the app what it needs to run for real:

- **A backend:** a Postgres database, server code, and jobs that run on a schedule, on a webhook, or when data changes.
- **Its own address**, behind your company sign-in (Entra, Okta or Google), for the people it is shared with.
- **A home on your company's own servers**, not anyone's laptop and not an AI vendor's.
- **Tools for every agent:** an app's actions are tools any agent can call through Harbor's MCP server, with each person's own access.

The agent is finished once it publishes; the app keeps working without it. Ordinary code does the routine work and calls a model only where judgment is needed, so nothing reruns a prompt every time the job runs.

Harbor runs apps written to its folder contract, not arbitrary programs. Apps are TypeScript and their screens are web pages. A tool someone already wrote in Python gets ported to TypeScript by the agent.

## For teams: an app is one folder

Your agent writes it. Harbor runs it. An app is one folder: a README, `app.ts`, and folders for its screens, actions, data, triggers, workflows, agents, skills and tests. Harbor checks it, publishes it and runs it for the people it is shared with.

### The parts of an app

- `README.md`: what it does, for people and agents.
- `app.ts`: its name, owner and connections.
- `ui/`: the screens your team opens. Optional: an app can be only tools, data and jobs.
- `actions/`: what people, agents and other apps can call.
- `data/`: tables, shared or personal.
- `triggers/`: what starts it: an event, a time, a webhook.
- `workflows/`: multi-step jobs that pick up where they stopped, and can wait for a person before carrying on.
- `agents/`: its own AI, with the tools and limits you set, run as the app.
- `skills/`: instructions that teach any agent to use the app.
- `tests/`: checked on every release. A release whose tests fail never goes live.

Here’s one: a deal desk, following one deal from start to finish.

- **Ask Claude for a discount approval app. It’s live today.** Claude writes the folder. Harbor tests it, then runs it at its own address behind your company’s sign-in.
- **Built in Claude. Used from ChatGPT.** Jared asks ChatGPT for 30% off Bachmanity. Any agent can call the desk’s actions through Harbor’s MCP server, so his request still goes to Monica in finance, whichever agent he asks.
- **The app’s own agent researches every request first.** Each new discount request starts the desk’s agent. It reads the customer’s Salesforce history and usage: margin after the discount 61%, paid on time for two years, usage up 40%. Recommend: approve.
- **Finance approves in Slack. The app updates Salesforce.** A workflow waited for Monica in finance to say yes, then wrote the new price to Salesforce.
- **A deal closes at the wrong price. The app catches it.** Bachmanity closed at 35%, not the approved 30%. The moment Salesforce says so, a trigger checks the price, and the desk flags it to Monica in finance.
- **Change the rule once. Every agent follows it.** Team leads can now approve up to 25%. The tests run on the new release before anyone uses it, and every approval so far stays.

Anything can be an app: tools for your own order system that any agent can use, a company brain every agent searches and adds to, partner payouts that wait for finance before paying in Stripe, a router that logs every email reaching the sales inbox and hands it to the right rep, or checks on every file a supplier sends before it reaches purchasing.

## For admins: one place for everything agents build

Today every agent connects to every system on its own. Harbor sits in between: connect each system once, and Claude, ChatGPT, Copilot, Cursor and Codex can all use it. Every app your agents build lives there too, with your skills. What one agent builds runs in Harbor, for the whole team, and you can share what’s in Harbor with your team: a skill Jared shared reaches Richard’s ChatGPT. Each person’s agent reaches what’s shared with them: Jared’s Claude and Richard’s ChatGPT both reach the shared deal desk and discount policy, and only Richard’s reaches the partner payouts. All of it in one place, in your own cloud, behind your sign-in.

## Your company’s systems and skills, in every agent

- **Connect your apps once, not once per agent.** Today every person connects their own Claude, ChatGPT and Cursor to Salesforce, Gmail and Slack. With Harbor, an admin connects each one once, and every agent connects to Harbor at one address. People still see only what they could see before, and every call is logged.
- **Agents can work in sites with no API, without your password.** Supplier portals and old internal sites need a browser. Harbor signs the agent in, and the agent never sees the password.
- **Write a process once. Every agent follows it.** Pricing rules, brand voice, refund policy: write them once as skills, and every agent follows them, instead of each person pasting them into chats.

## Not an artifact. Not a Site.

They can look alike. They’re built different.

Claude artifacts and ChatGPT Sites are:

- Pages inside Claude or ChatGPT.
- People open them and use them.
- For people on that AI plan, or the public.
- Run on Anthropic’s or OpenAI’s servers.

Harbor apps are:

- Software of their own, outside any AI app.
- People, agents and other systems all use them. Their actions are tools any agent can call.
- For everyone at your company, whatever AI they use, or none.
- Run in your own cloud, next to your internal systems.


### In detail

The two can look similar: a Claude artifact or a ChatGPT Site is a page an AI built that people can open. A Harbor app can be a page too, but it is software of its own: it can also be data a team shares, tools agents call, or jobs started by a schedule, a webhook or an email, and anyone at the company can use it, whatever AI they use.

| | Claude artifact | ChatGPT Site (public beta) | Harbor app |
| --- | --- | --- | --- |
| What it is | One page, with no backend | A website with server code, a 10 GB database, file storage and secrets | A whole app: screens, server code, data, triggers, workflows and agents |
| Repeating work | A separate scheduled task or routine: an AI agent reruns a prompt, at most hourly | A linked ChatGPT scheduled task: an AI agent run, at most hourly, owned by one person | The app's own code, run as the app |
| Other agents can use it | No | No (ChatGPT's own desktop browser can use its page tools while it's open) | Any agent over MCP |
| Who can open it | Members of your Claude organization | Members of your ChatGPT workspace, invited viewers with accounts, or anyone if public | Anyone in your company directory you share it with; no AI seat needed |
| Who decides access | Claude's sharing settings | Site sharing, then rules you write in the site's code | Harbor, outside the app's code: sharing, data scopes and admin rules per tool |
| Where it runs | Anthropic's servers | OpenAI's servers; private networks aren't supported | Your own cloud or your own machine |
| Storage | 20 MB of text per artifact | 10 GB database, plus files | Postgres |

What a ChatGPT Site does have, and Harbor doesn't claim otherwise: server code, a database, files, secrets, calls to outside services, workspace sign-in, each visitor's own connected apps, saved versions and ownership transfer.

**Microsoft Copilot Managed Runtime** (public preview since September 2026) is the closest product to Harbor from an AI vendor. Apps built in Copilot Cowork, Copilot Studio, or by a coding agent with Microsoft's CLI are hosted by Microsoft inside the Microsoft 365 tenant, with Entra sign-in, Microsoft's connectors, data loss prevention, sharing and an admin inventory. Its documentation does not describe a built-in database, background or scheduled work, test runs before deploy, or other agents calling an app. Apps reach data through connectors as the signed-in user, every user needs a Power Apps Premium license or pays Copilot Credits per call, and it runs only on Microsoft's hosting with Entra sign-in.

**Scheduled tasks, routines and workspace agents** (Claude, ChatGPT, Claude in Slack) are good for a recurring job an agent can do from instructions. Each run is a new AI agent session, owned by a person or a channel, at most hourly. A Harbor app runs its own code and can call a model where it needs one.

**Hosting it yourself** on Vercel, Replit or a cloud VM gets the code online, but each app then needs its own sign-in, database, schedules, secrets, access rules and agent tools, and IT can't see it. Harbor gives every app those once, in one place your admins see.

### When you don't need Harbor

- **A question or a one-off job:** a chat answer or a script is enough.
- **A single page someone opens**, like a calculator, a form or a dashboard over your own connected data: a Claude artifact, a ChatGPT Site or an app inside the chat works well.
- **A recurring job an agent can do from instructions:** a scheduled task or routine works.

Harbor matters once a tool has to be more than a page, be used by people and agents across different AI tools, or run in your own cloud.

Checked against Anthropic’s, OpenAI’s and Microsoft’s documentation on 17 October 2026. These products change quickly.

## Runs in your cloud, behind your login

Harbor is open source. You install it in your own cloud account, or on one computer just for you, so your apps and their data stay with you. Everyone signs in with the login your company already uses: Entra, Okta or Google. Admins see every app and who uses it. Apps never see a password or API key, and every action is logged.

## Questions

**What is a Harbor app?** Software your agent builds for your team: an approval tool, a shared tracker, a report that runs every morning. Your agent writes the code. Harbor runs it in your cloud, behind your company login, for the people and agents you share it with.

**What is the Harbor gateway?** The one place your agents reach everything at work. You connect your tools to Harbor once: Salesforce, Gmail, Slack, an internal API. Your skills, your apps and resources like a signed-in browser live there too. Then you connect each agent to Harbor’s one MCP address, and it can use all of it, with each person’s own access. An app’s actions show up as tools any agent can call. The browser signs agents in to sites with no API, and they never see the password.

**Does everyone need a Claude or ChatGPT seat?** No. Anyone in your company directory can open an app shared with them. Agents are only needed to build and change apps, and to use them from chat.

**Which agents work with it?** Any agent that speaks MCP: Claude, ChatGPT, Copilot, Cursor, Codex and others. An app built in one can be used from all of them.

**How is a Harbor app different from a Claude artifact or a ChatGPT Site?** An artifact or Site is a page inside one AI product. A Harbor app is software of its own: a page, a database your team shares, tools any agent calls, or jobs that run on a schedule or when something happens, for everyone at your company.

**How do people sign in?** With your company login: Entra, Okta or Google. Admins see every app and who uses it, and app code never sees a credential.

**Where does the data live?** In Postgres, in your own cloud or on your own machine, next to your internal systems.

**What happens when an agent changes an app?** Harbor installs, type-checks, builds and runs the app’s tests. A release that fails never goes live, and the app’s data stays.

**Is Harbor open source?** Yes. It’s in private preview now. Join the waitlist and we’ll invite your company.

## Contact

charlieconner04@gmail.com