Know what Hatch can build.
Then get the best out of it.

01 Start here

What Hatch is—and what it is not

Hatch is a desktop app builder for small, self-contained tools tailored to one person or one team's real workflow. It plans the app with you, generates it, runs automated checks, installs it locally and keeps a version history for later changes.

02 Choose the right job

Where Hatch works best

A good Hatch app has a clear user, a repeatable job, a bounded set of records and an observable result. If the value depends on a live online service, multiple simultaneous users or unattended automation, Hatch is the wrong foundation in its current form.

03 Build flow

What happens after you describe an app

1. 01

   Describe the job

   Say who uses it, what they do, what the app remembers and what it should produce.

2. 02

   Answer focused questions

   Hatch resolves missing fields, rules, terminology and decisions that would materially change the result.

3. 03

   Review the blueprint

   You see the workflow, data and a sketch before generation. Treat this as the contract; mistakes here flow into the app.

4. 04

   Generate and check

   Hatch writes the app, validates it and smoke-launches it with both empty and seeded data. Failed checks can trigger repairs.

5. 05

   Visual review

   A strong vision-capable model judges the running app against the blueprint and can request corrective passes.

6. 06

   Install, test and refine

   Open the shortcut, use realistic records and request precise changes. Versions are append-only, so reverting creates a safe new version.

04 Scope before speed

Pick the smallest build depth that fits

The one-minute, roughly twenty-cent example on the home page is a measured small app—not a promise for every scope. Complete and Ambitious builds can run for hours and use substantially more model tokens. A narrow version 1 is usually easier to verify and cheaper to refine than an oversized first attempt.

05 Prompt like the operator

How to get the strongest result

Do not describe a software category. Describe the work as the person doing it: the inputs, rules, sequence, exceptions and final output.

I run a residential cleaning business and need a quote builder for one person.

A quote records the client's name, property size, visit frequency, base service and optional extras. Price equals the base rate by bedroom count plus selected extras, with 10% off weekly visits.

I need a clear quote summary I can copy and export as a PDF, plus a searchable history of past quotes. I usually have fewer than 500 quotes.

No online payments, customer login or cloud sync.

During questions and refinements

• Answer decisions that change the workflow; use You decide for low-stakes details.
• Read every blueprint section, especially stored data, calculations and “not included.”
• After the build, test with a normal case, an empty case and a difficult edge case.
• Request one precise behavioral change at a time: what you did, what happened and what should happen instead.
• Use real terms from the business. The right labels make the app feel tailor-made even when the workflow is simple.

06 Give useful references

Attachments Hatch can use

You can attach up to five images or text references while describing an app. High-value attachments include a screenshot of an existing tool, a photo of a paper form, a sketch of the desired screen, a sample CSV and a short rules document.

Accepted

Images, CSV, TXT, Markdown and JSON

Limit

5 files

Text read

First 20,000 characters per file

Not accepted

PDF and Excel .xlsx references

07 Model guide

Use the strongest model where judgment matters

Hatch divides the work into roles. This is why a mixed setup often beats paying for the most expensive model on every token.

Current provider guidance

Model availability changes. Refresh the provider list before pinning. Anthropic describes Fable 5 as its highest-capability model, Opus 5 for complex coding and Sonnet 5 as the quality/speed balance. OpenAI's current guide positions GPT-5.6 Sol for complex reasoning and coding, Terra for intelligence/cost balance and Luna for cost-sensitive volume. Kimi K3 is a multimodal coding model with long-context and visual iteration strengths. Hatch's other OpenRouter stage choices are implementation defaults from the current app source. See the first-party model sources.

08 Hard capability boundary

What a generated app can and cannot do

09 Know where data goes

Data and privacy

A finished app without AI permission works offline. An app with AI permission needs the connected cloud provider or local model server; its runtime prompt is sent there. Using local models keeps generation local, subject to the behavior and configuration of that server.

10 Keep cost in the formula

How model costs work

Hatch is free and adds no markup. Cloud API usage is billed directly by the provider behind your key. A ChatGPT or Claude subscription does not automatically include API usage.

• The home-page example used $0.16 in model calls on Balanced. It is a measured small build, not a universal average.
• Scope, model prices, retries, attachment size and later refinements all affect total cost.
• Balanced is the default value choice. Cheapest trades planning/build quality for cost; Best applies the strongest model throughout.
• Local models have no per-token provider bill, but use your hardware and can take much longer.
• Hatch tracks reported monthly spend and shows a guardrail, set to $20 by default. Treat it as tracking—not a provider-side hard stop.
• Set a real spending limit with the provider. Runtime AI inside each generated app has its own Hatch-enforced budget, $2 by default.

11 Two export formats

Share with or without Hatch

A .hatchapp import shows the author, description and requested permissions before installation and runs static validation. Current exports do not include the app's saved records. Standalone apps may also trigger unsigned-app warnings from Windows or macOS.

12 Sandboxed, then reviewed

Safety model

Generated apps run with Node.js integration off, context isolation on and a restrictive content policy. They receive only declared capabilities such as local storage, file pickers, notifications or AI. Hatch scans generated files, launches the app in hidden empty and seeded test states, and uses a visual reviewer before presenting it as ready.

13 Current release

Known limitations worth knowing up front

• No direct Google provider. Current connections are Anthropic, OpenAI, OpenRouter and OpenAI-compatible endpoints. A compatible gateway may expose other models.
• No internet except AI. Generated apps cannot fetch APIs, scrape pages, use webhooks or silently connect elsewhere.
• No background or scheduled work. Notifications work only while the app is open.
• No cloud sync or live collaboration. Saved app data is local to one computer.
• Limited build references. Up to five images/CSV/text/Markdown/JSON files; no PDF or Excel reference import.
• Imported apps are not full projects. A shared package does not restore the original build conversation for editing.
• No hosted gallery. Sharing is file-based.
• Unsigned installers and exports. Windows SmartScreen and macOS may display warnings.
• Model access varies. Your provider account, credits, regional availability and rate limits still apply.
• Quality varies. Long or underspecified apps can exceed a model's practical reasoning or output reliability even when token limits are sufficient.

14 When a build struggles

Common fixes

15 Before you rely on it

A short final checklist

1. 01

   The app has one clear user and a bounded primary job.

2. 02

   The blueprint names the right fields, terminology, calculations and exclusions.

3. 03

   The selected Blueprint model supports images and is strong enough for the scope.

4. 04

   You tested normal, empty and edge-case records with sample data.

5. 05

   You verified calculations, exports, deletion and recovery yourself.

6. 06

   You know whether build material or runtime AI prompts go to a cloud provider.

7. 07

   The provider—not only Hatch—has an appropriate spending limit.

8. 08

   The app is not being treated as independently verified high-risk software.