API Maker
The backend engine for the AI era
SandboxesEpisode 46 6:36 10 chapters November 2026
Your code runs in eight languages, each in an image built with the packages it needs, apart from the server and from other accounts. Tour Sandbox Settings: Node.js or Bun, packages from PyPI, npm and NuGet, a Dockerfile of your own with guard rails, and where code runs: lean, sandbox or native.
In the playlists The complete course Run in production
Your code runs in eight languages. Each language runs in an image, built with the packages it needs. It is kept apart from the server, and from other accounts. In this video, we look at sandboxes, runtimes, and packages.
A runtime for each language usually means work. You write a Dockerfile for each service, and patch it by hand. Packages are installed on one machine, and missing on the next. And code downloads what it wants while it runs. With API Maker, one page sets the runtime and the packages of each language. Every server builds the same image, named by what it holds. And only the listed packages exist where the code runs.
In this video: Sandbox Settings, a PyPI package in a Python API, your own Dockerfile and its guard rails, where code runs, and the sandboxes of a server.
From the dashboard, open Sandbox Settings. There are eight languages, each with the image it runs in, ready on this server. JavaScript has two runtimes: Bun, and Node.js, the default. A piece of code can name its runtime. The list holds the released versions of its image, from the oldest one API Maker runs. None is picked, so it uses the default. Python has its slim image, and no package: just the standard library.
Open the Dockerfile. API Maker fills in the image, the packages, and the runner. This is what is built. Then the tab Images and sandboxes shows the image on every server, ready.
In API Testing, pick Sales Summary, a Python API that imports humanize. Send. We get five hundred: no module named humanize, at line one. Nothing is downloaded while code runs. Open Operations, then Sandbox Settings, and Python. Type humanize, the exact name that PyPI knows, and you see its versions, from PyPI. Pick the latest, click Add, and Save. The image of Python is built on every server, and now it is ready.
Open the menu, then API Testing, Custom APIs, and Sales Summary. Send. We get a hundred and eighty orders, and their total in words. The code imports humanize, like any Python package. The orders come from MySQL, and humanize writes the total for a person to read. Every language has its own package store: npm, PyPI, Packagist, NuGet, Go modules, crates.io, and Maven Central.
Now your own Dockerfile. Customize shows the default one, with its three markers. We add our own ENTRYPOINT line, and Final Dockerfile refuses it, with the reason. The entrypoint belongs to API Maker, because it starts the program that runs your code. It also refuses a second marker, or a file without FROM. Click Reset to default, turn Customize off, and Refresh. Nothing was saved.
Code runs in one of three places. TypeScript that only reads runs inside the server itself, in a few milliseconds. Everything else runs in a Docker sandbox of its account and language. It is a container, with an agent inside that answers g.sys calls. Code marked runOnNativeProcess runs in a runner on the server machine, shared by every account, for tools the machine has. And one API can have its own sandbox group, with its own runtime and packages.
Pick Python, and in the tab Images and sandboxes, you see its sandbox on this server, with its runners. The sandbox count is the most sandboxes each language can have on each server. A second one starts while the first is busy. Add one more, to three. We do not save it here. Open the menu of the account, click Restart All Sandbox, and Yes. Ask again, and there is no sandbox.
The next call starts one again. Refresh, and the count stays at two. In API Testing, pick Custom APIs, then Sales Summary. In Header, set x-am-meta to true, and Send. The meta of the answer says where it ran: the server video1, the port of its sandbox, and the runtime python. Only the admin panel gets this. Another caller gets the answer without it.
Good to know. A package can be switched off. It stays in the list, but it is not installed. Accounts with the same settings share one image, and images that nothing asks for are removed every hour. The Vulnerabilities page checks exactly these packages.
Each language has one list of packages, and nothing is downloaded while code runs. Every server has the same image, named by what it holds. So if an AI assistant writes code with a package that is not in the list, it fails at its line, at once, instead of pulling it in.
To sum up. Eight languages, with their runtimes and versions, are on one page. Packages come from the store of each language. You can use your own Dockerfile, with guard rails. Code runs inside the server, in sandboxes, or in native runners. And sandboxes restart in one click. Next: deploying a new version of API Maker. API Maker. Build faster, spend smarter, deliver better.
Length 6:42Deploy
Length 8:36Heavy APIs
Length 6:16Per admin
Length 5:53One call