The honest answer to "does it integrate?"
Every piece of business software claims to integrate with everything. Look closely and the claim usually means one of two things: there is a list of logos on a page, or there is an API and you are welcome to build something against it.
Both are true. Neither helps on a Tuesday afternoon when you want a competitor brief to land in Slack and you are not a developer.
The gap between "it is possible" and "it is working" is where most integrations die. Not because the technology is hard, but because somebody has to sit down, read two sets of documentation, and connect them — and that person is usually you, and you have other work.
App Hub takes a different route. It does not ship a workflow builder for you to learn. It assumes something else: that you may already pay for an AI assistant that can read, click and type in your own browser.
If you do, you do not need a workflow builder. You have an integrations manager.
What actually changes
Claude Code is an AI assistant that runs on your machine. With the Claude in Chrome extension, it can act on pages in your browser rather than only read files — which means it can do the setup instead of describing it to you.
That distinction matters more than it sounds. Advice about integrations is cheap and everywhere. Somebody sitting at your screen, opening the right settings pages in the right order, and telling you exactly which value goes where, is the part nobody provides.
The starter pack
App Hub publishes a small download: the integration starter pack. It is a folder of instructions written for the agent, not for you.
Unpack it, point Claude Code at it, and it learns three things: what App Hub is, how it connects in both directions, and how to behave while doing it. From there you describe the outcome you want in plain English.
Not "configure a webhook". Something more like:
"I want a competitor brief posted into our market-intel Slack channel whenever I run one."
Or:
"When a deal in our CRM reaches Proposal, research that company and write the brief back to the deal record."
Those are specifications. The agent turns them into working plumbing.
It goes both ways, and that is the interesting part
Most tools can send you a notification. Fewer can be asked to do something by another system.
Outbound is the familiar direction: App Hub finishes a piece of research and pushes it somewhere — Slack, email, Telegram, or a webhook pointed at Zapier, Make or n8n. Point it at Zapier and one result can fan out to everything Zapier already connects to, without App Hub needing to know any of those tools exist.
Inbound is the direction that changes what the product is. App Hub can be called. Something else — a CRM, a form, a schedule — asks it to run, and it runs, with the same limits and protections as if you had clicked the button yourself.
Once both directions exist, App Hub stops being a place you visit and becomes a step in a sequence. Grounded research with its sources attached, arriving where the work already happens.
The overnight research desk
Here is where it stops being about integrations and starts being about time.
Because the schedule lives with your Claude Code rather than inside App Hub, a scheduled job is not a fixed timer firing the same request forever. It is an agent turn. It can decide things.
So an instruction like this is reasonable:
"Every hour tonight, research a different one of these eight topics. Turn each into a blog piece in the Content Studio. At the end, push them all to my Slack channel so I can read them on the way to work."
Eight pieces of grounded research, transmuted into eight drafts, delivered in one batch before breakfast. Nobody opened the app. And the last instruction — push them to Slack when you are done — is a sentence in a brief, not a workflow you configured.
That is the difference between an automation platform and an agent. A platform needs the sequence drawn out in advance. An agent needs it described.
What it will not do, on purpose
Two limits are built into the starter pack deliberately, and both are in your interest.
It will not handle your credentials. The agent will navigate to the right screen and tell you precisely which value belongs in which field, then hand back to you. API keys, webhook URLs and passwords are typed by you. An assistant that hoovers up your secrets to save you ten seconds is not saving you anything worth having.
It will not run analyses to explore. Runs happen on your own AI provider key and cost you money. The agent is instructed to say what a step costs before taking it, and to test with one real run rather than several speculative ones.
There is a third limit worth knowing. If something fails, the agent may retry it once — then it stops, writes down what happened, and tells you. Not twice, no clever escalating loop. An automation that keeps trying while you sleep is how a small problem becomes an invoice.
Keeping an eye on it
Anything running unattended should leave a trail. When the agent sets up a schedule it will offer to write one row to a spreadsheet at the end of every task: what ran, whether it worked, how confident the research was, how many sources it found, anything it had to fix.
That log is not paperwork. It is the difference between knowing your overnight desk is working and assuming it is. A job that has been quietly failing for three weeks is worse than no job at all, because you believed you were covered.
Getting started
You need three things: Claude Code, the Claude in Chrome extension, and an App Hub account. Download the starter pack, open Claude Code where you unpacked it, and describe what you want.
Begin with something small and outbound — a result landing in Slack. It needs no API key, it takes a few minutes, and it puts something in front of your colleagues that they will actually notice. Build the overnight desk after you have watched the simple version work.
The reason this works is not that App Hub is clever about integrations. It is that App Hub deliberately does not try to be. It stays a research function that can be called and can call out, and lets the agent you already pay for handle the wiring, the schedule and the recovery. Fewer features to learn, and a ceiling set by what you can describe rather than by what somebody built into a menu.
Ofte stilte spørsmål
No, but you do need Claude Code and the Claude in Chrome extension. The pack is written for the agent rather than for you — your part is describing the outcome you want in plain English and typing your own credentials when prompted. The agent does the clicking.
No, and deliberately so. Scheduling lives with your own Claude Code or cron. That is more capable than a built-in timer, because a scheduled agent can read each result and decide what to do next rather than repeating one fixed request.
The agent is instructed to retry once at most, then stop, log the failure and notify you. It never retries errors caused by account state, such as a missing AI provider key or an exhausted allowance, because repeating those on a schedule turns one problem into a nightly one.