Frequently asked questions

Cryogram turns a job you describe into a working automation. You explain what you want in a chat, the AI figures out how to do the job once in front of you, then writes what it did into ordinary code. Code gives the same output for the same input every time, and AI is used only in the steps that need judgement, such as scoring a CV or drafting a reply. What you end up with is a button: press it, the job is done, and every run is checked and logged, all on your own machine.

An agent thinks the whole task through on every run. That makes each run slow, it uses your tokens or quota, and the output can change from one day to the next, so an agent works well for one-off tasks and poorly for a task you repeat. Cryogram uses AI once, while building, and writes the result into code: the routine steps run identically every time, a result that fails a step's check is never passed on to the next step, and a run whose steps use no AI costs nothing.

The script is the smallest part. To trust one at work, you also need a safe place for passwords and keys, a contained place to run it so a mistake doesn't spread, checks on each step, a clean stop when something fails, and a record of what ran and what came out. Setting that up takes far longer than the script itself, and until it all works, you are the one debugging the drafts. Cryogram has all of it built in, and it tests every step by doing the job on your data before the step is saved, so the automation you get has already done the job correctly.

File pipelines, document intake and reading, email digests, spreadsheet-to-report flows, CRM updates, web checks, outreach merges - most repeatable tasks that mix fetching, reshaping and a few judgement calls. Because it runs on your machine, it reaches what cloud tools cannot: local folders, network shares, internal systems, and sites with no API. The examples on the front page show what early users have built.

Cryogram runs on macOS and Windows.

Building runs on a Claude or ChatGPT subscription, or an Anthropic or OpenAI API key. The steps inside an automation can use Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, xAI, Moonshot or OpenRouter out of the box, and you can add any provider with an OpenAI-style endpoint, local models included. Everything runs on your own keys or subscriptions, and we never resell or mark up AI usage.

No. Everything the app needs comes with it, including the skills - the knowledge the AI uses when it builds - so it works offline from the first start. It can check for newer skills and app versions when it starts, and both checks just read a published file without sending anything about you; each has its own switch, and app updates never install without your approval. Building workflows still needs your AI provider, which is an internet service.

Yes, and they need to learn nothing. You share the automation, and your colleague gets a button and an input form: they press Run, watch the progress in plain language, answer when the automation asks them something, and get the results, with a full history of earlier runs beside them.

The automation stops mid-run rather than guessing, and a result that failed a check is never passed on to the next step. You get a plain-language account of what stopped and why. The AI then investigates what changed and proposes a fix, and the fix is applied only after it passes the automation's checks against the recorded results of past good runs. Anything that touches an outside system asks you first.

No, and you should be wary of anyone who does. Our claim is that automations are dependable and honest: every step is tested, every run is checked, and when the world changes - a website moves, a file format shifts - the automation stops and tells you. Then it gets fixed, and the fix is tested before it is applied.

Any step that acts on an outside system - sending an email, writing to a CRM - asks for your approval first. For each automation you choose whether those steps need your confirmation on every run or are approved to run unattended. The security page describes the full picture.

On your machine and in your own systems. Automations read inputs and write outputs at run time and keep none of your business data in between, and the material from building - your chat, sample files, recorded test evidence - never leaves your machine. The full picture is on the security page.

No. The AI sees secret names only. The values are stored encrypted on your machine, filled in at the moment a step runs, then removed from outputs and logs. There is no code path that hands a secret value to a model.

Nothing, and there is no account or identity to send it under. The one thing that can leave your machine is a problem report: when a run stops, the app can show you a report about the failure, and it goes only if you read it and press send - never your data, your conversations or a credential, and never on its own.

We are opening access in batches, so that everyone who joins gets properly set up and supported. Join the waitlist and we will email you when your access is ready. Telling us what you would like to automate is optional, and helps us send you the right getting-started guidance.

Very likely. The first automations tend to be your own: the report you build every Monday, the inbox you sort by hand, the list you keep updated. Once one of them works well, you can share it with a colleague, who runs it without learning anything.