Skip to content
Blog

One API key, your whole cloud

Why Harakumo exists: shipping a small product should not require assembling six companies. And why the whole platform speaks MCP, so an AI assistant can operate it end to end.

2 min read

  • Launch
  • AI connector
On this page
  1. Deploys that figure it out
  2. The part that is actually new

Every product I shipped this year started the same way: three days of wiring before the first line of the actual product. A database somewhere, auth somewhere else, storage, email, payments, a domain — six signups, six SDKs, six bills, and an environment file nobody could decode a month later.

Harakumo is that wiring, extracted and made into the product. Hosting and deploys, edge SQL, object storage, vector search, memory, functions, auth for your app's own users, payments, transactional email, domains, media, an AI gateway across the major model vendors, agents, CDN caching, country blocking, analytics and audit logs — behind one API key, one dashboard, one account.

Deploys that figure it out

Connect a GitHub repo or upload a folder. Harakumo works out what it is — a static site, a server app, an edge program — and deploys it accordingly. There is no build configuration to write, because classification is the platform's job, not yours. Finished output goes live directly; source that needs building runs on build machines with a package cache that carries across deploys.

The part that is actually new

The whole platform is exposed over the Model Context Protocol. You add Harakumo to Claude — other MCP clients speak the same protocol, though only Claude is verified so far — sign in, and the assistant gets the same tools you have: create the database, provision storage, deploy the app, buy the domain, read the logs.

Two design rules make that safe. The connector acts with your role — it can do what you can do and nothing more, and revoking it takes effect on its next call. And every change it makes to your setup lands in your audit log like any other, while everyday calls (a query, an email, a checkout link) are counted as usage, so "what did the AI change" is never a mystery.

Most platforms are bolting an AI layer onto a product shaped like a dashboard, which is why those integrations can read but rarely act. We built the dashboard, the API and the agent surface as one thing. That is the bet: the console stops being the primary interface, and infrastructure that agents can fully operate wins.

It is live now, free to start, at harakumo.com. Deploy something and tell me where it breaks — that is genuinely the most useful thing you can do for us this month.

Filed under

  • Launch
  • AI connector

Go deeper in the docs

Quickstart: your first app

From signup to a live app with a database, browser uploads to storage and a payment link, on the free Hobby plan, with code that runs as written: one worker.js file, no server to run.

AI connector (MCP)

Connect Claude, ChatGPT or any MCP client to https://harakumo.com/mcp and run the platform from a conversation, read-only or with permission to make changes.

Keep reading

2 min read

A failed deploy should never take your site down

We found the worst kind of bug in our deploy pipeline: a build that failed could take the running app down with it. Here is what was wrong, and the property that fixes it for good.

  • Deploys
  • Reliability

2 min read

Errors that tell the truth

A platform earns trust in its failure messages, not its success screens. The rules we follow when something goes wrong: name the limit, name the state of your money, and never pretend.

  • Errors
  • Billing
  • Domains

Build something real today.

Start free with no card. Connect your AI and ask it to build, or deploy a folder yourself.