Sign in with a passkey through your own browser.

You updated Hatch, and this is what changed.

A site that asks for a passkey now opens in Chrome, Arc, Brave or Edge, where Touch ID and your phone work, and Hatch brings the sign-in back.

Passkey sign-ins

When a site such as Apple or Microsoft asks for a passkey, Hatch offers to open it in your own browser. Sign in there, press Bring my sign-in across, and the Hatch reopens the page signed in.

Any sign-in, from the menu

Right-click a Hatch and choose Sign in through another browser to do the same for any site that refuses to sign you in inside Hatch.

More icons for action buttons

Action buttons now offer 54 icons, with search, terminal, database, news, wallet, inbox and more joining the set.

Page sizes your agent can trust

When an agent resizes a page, Hatch checks that the page followed and tells the agent when it has not.

Put your agent to work

The whole guide follows below, so scroll on if you want the tour again.

Thanks for trying out Hatch.

The browser your agent drives while you watch.

A Mac browser for people who build with agents.

Your agent writes the code and you judge the result. Both of you need the live site in front of you. Hatch puts it there, between the Paper or Figma file and the finished build.

Get Hatch

Download Hatch for Mac

Hatch needs a Mac with Apple silicon.

Or paste this line into Terminal. It puts Hatch in Applications, opens it, and macOS asks no questions.

curl -fsSL https://raw.githubusercontent.com/jamiejefferson/hatch/main/install.sh | sh

The download asks one extra step the first time. Hatch has no Apple Developer ID yet, so macOS says it cannot check the app. Choose Done, open System Settings, then Privacy & Security, and choose Open Anyway beside Hatch.

This page runs to eleven stages, and most of them give you a thing to try.

The eleven stages below tour Hatch. Open this page in Hatch and each stage gives you a thing to try.

DesignPaper and Figma draw the site.

HatchHatch holds your prototypes.

BuildCode and deploy finish it.

We left out everything you never use.

Hatch runs on the engine inside Chrome, so every site works. It has no address bar and no extensions. We added one thing. The agent view turns a page into a short outline with a reference on every line, and your agent acts on a line by its reference.

Regular view

Agent view

[e12] region "2 · The browser"
[e13] heading "We left out everything you never use."
[e14] text "Hatch runs on the engine inside Chrome…"
[e15] button "Done it!"

Press these in Hatch

The view switchIt sits in the toolbar beside the canvas. The eye shows the page and the outline shows the agent view.

Try it

Select this Hatch and press the view switch in the toolbar. Find this sentence in the agent view, then switch back.

Now imagine your pages side by side.

Hatch has kept something from you.
Press Esc and see.

This is the canvas. Every page on it is a Hatch.

Pan with two fingers and pinch to zoom. Select a Hatch to scroll its page. Now double tap the empty canvas and open a second Hatch beside this one.

Fit to view fills the window with one page.

The canvas holds as many pages as you open.

Press these in Hatch

Fit to viewThe bar above a selected Hatch holds it. Esc leaves it.
New HatchThe plus beside Hatches in the left column opens it, and a double tap on empty canvas does the same.

Go back any time

The Fit to view button above this page fills the window again, and Esc brings the canvas back.

Put your comment on the thing you want changed.

A comment in Hatch pins to an element on the live page and stays there. Come back next week and you will find it where you left it. Your agent reads it, makes the change and tells you what it did.

You can resolve a thread or edit your words. Every comment is also a plain text file.

A heading that runs a little longer than it should

A line of supporting copy sits under the heading.

Press these in Hatch

HatchSelect this Hatch.
→
Add a commentThe button sits in the toolbar beside the canvas. Press it, then press the card.

Try it

Select this Hatch. Press the comment tool in the toolbar and pin “Make this heading shorter” to the card. The Comments tab in the sidebar lists every thread.

See your page on a phone and a desktop at the same moment.

Five buttons set a Hatch to desktop, laptop, tablet upright, tablet on its side or mobile. Open one page in two Hatches and you catch the broken layout before your client does.

1440 × 900

This Hatch is at the Laptop size.

The five size buttons sit under Size in the Hatch tab of the sidebar

Desktop960 × 752
Laptop1440 × 900
Tablet portrait768 × 1024
Tablet landscape1024 × 768
Mobile390 × 844

Try it

Choose Mobile in the Hatch tab and watch this page reflow. Then open it again at Desktop, side by side.

Every project gets its own address, and keeps it.

Give Hatch a folder and it starts the dev server for you. The project answers at a short name such as hatch:my-site, on a port that never changes. You will not open localhost:3000 and find the wrong site again. Hatch also lists the servers you started elsewhere.

Start or stop each server from one panel, and read its logs there.

~/Sites/my-sitehatch:my-site4300Running

~/Sites/shophatch:shop4301Stopped

~/Sites/bloghatch:blog3000Found

Press these in Hatch

ProjectsOpen Library in the left column.
→
Add a projectHatch asks you for a folder.
→
OpenHatch opens the project in a new Hatch.

Try it

Open Projects and add a folder. Then open it in a Hatch beside this one.

Hand Hatch to your agent.

One button in Settings copies the details. Paste them to your agent and it sets itself up. It works on its project’s own canvas, so your view stays as you left it. The Activity panel records every call it makes.

Hatch works with any agent that speaks MCP, and we test it with Claude Code.

Press these in Hatch

SettingsThe sliders icon at the top right opens it in the sidebar.
→
Copy the details for your agentPaste what it copies to your agent.
ActivityThis panel lists every call your agent makes.

Try it. Send your agent this.

Your agent needs the details from Settings first.

Open the Hatch guide, press the button in stage 7, then read my comment in stage 4 and reply to it.

The page answers when it feels the press.

After the press

Open the comment you pinned in stage 4 and read your agent’s reply.

OptionalWorks with Jev

Hatch works with Jev from TypeSafe.ai.

36/36

Jev picked the right element every time in our test, in a median of 0.8 seconds.

Jev is a small model built to make fast, inexpensive, structured decisions that software can use directly. Hatch asks it which element your agent means, so your agent says “the button that refuses cookies” and Hatch presses it.

Your agent can also hand Jev a whole goal, such as “search for "blue badge" and open the first result”. Jev chooses each click and keystroke, Hatch carries it out, and your agent reads a record of every step before it carries on. Your agent keeps the decisions and Jev takes the repetition. Stage 11 shows how a run works.

By hand

snapshotclicksnapshotfillsnapshotpress_keysnapshotclickone turn of your agent for every call

With Jev

jev_run {goal}one turn of your agent for the whole run

Connect it in Settings with your own key, from console.typesafe.ai/settings/keys or from openrouter.ai/keys.

Settings“Connect Jev” sits near the foot.
→
Save and connectPaste your key, then press this.

Jev stays off until you connect it, and “Disconnect Jev” switches it off again.
Each use sends the page’s outline to TypeSafe, through OpenRouter when the key is an OpenRouter key. A page where Hatch filled a sign-in is never sent.
The Activity panel lists each run with its steps, Jev’s confidence in each one and what the run cost.
Below 80% certainty Hatch does nothing and lists the closest elements.
To try it, connect Jev, then ask your agent to press “the sample button in stage 4”.

Take any element on the web into Paper.

“Grab an element” copies any element from any page. Paste it into Paper and it arrives as editable layers. Your agent can measure an element and read its styles. It can save the images too.

FrameSample card
TextHeading
TextSupporting copy
FrameSample button

Press these in Hatch

HatchSelect a Hatch.
→
Grab an elementThe button sits in the toolbar beside the canvas. Press it, then press the element.
Let agents run script in pagesYour agent needs this switch in Settings before it grabs anything.

Try it

Press “Grab an element”, pick the card in stage 4 and paste it into Paper. If you have no Paper, ask your agent for the card’s size and colours.

Your agent never sees your password.

Hatch saves a sign-in and asks you before it fills one for an agent. The password goes into the form and nowhere else. Google refuses sign-ins inside apps like Hatch, so Hatch brings yours across from Chrome, Arc, Brave or Edge.

macOS asks for your password when Hatch brings a sign-in across, because the other browser keeps its key in the Keychain.

Press these in Hatch

Sign-insSettings, at the top right, holds them.
→
Save this sign-inFill in the site, the username and the password first.

Try it

Save a made-up sign-in for this page under Sign-ins in Settings. Ask your agent to sign in here, and answer the question Hatch asks you.

Set up these four things once.

Links

Save the pages you open most. This guide is already there.

Canvas 1+

Canvases

Each canvas holds its own Hatches. Give it a project folder and the agents working in that folder use it.

Hatch Home

Choose the page every new Hatch opens. Today it is this guide.

Default browser

One button makes Hatch your Mac’s browser. Links from other apps open in a new Hatch.

Try it

Save a page you open every day in Links. Then right-click this Hatch and copy its link for your agent.

Let Jev take the repetitive steps.

A search, a long list or a checkout asks your agent for the same small decision again and again: read the page, choose the next click, wait, read the page again. Each of those is a full turn of a large model. With Jev connected, your agent states the goal once and Jev chooses each step.

Hatch asks Jev three things at every step: which action comes next, whether the goal is reached, and whether the run is stuck. Jev answers all three in one reply and writes no text, which keeps each step short. TypeSafe charges $0.042 for a million tokens read, and every run reports what it cost.

4/4

We gave Jev four goals on four public sites and it finished all four. On GOV.UK it refused the cookies, searched for “blue badge” and opened the right result: five steps in 3.0 seconds, for $0.0012.

Your agent sets the goal

“Search for "espresso machine" and open the first result under £300.” Any text Jev must type sits inside quotes, because Jev chooses and never writes.

Jev takes the steps

It clicks, types, chooses from dropdowns, presses Enter and scrolls. It stops when the goal is reached, when it is stuck, or when it runs out of steps or time.

Your agent reads the record

Every step comes back with Jev’s confidence and what the page did, beside the final page as an outline. Your agent checks the result and decides what happens next.

You see what it cost

The Activity panel shows each run as one line, such as “done (8 steps, $0.0031, 4.2s)”, and opens to show every step.

Jev suits mechanical goals. Your agent still does the reasoning: which product to pick, whether a result is right, what to try when a page surprises it.

To switch it on, open Settings, paste your TypeSafe or OpenRouter key under “Connect Jev” and press “Save and connect”. Then point your agent at the Agent page, which tells it when a goal suits Jev.

Help your agent get the most from Hatch.

We wrote a page for your agent to read. Point your agent at it once, and it learns how Hatch works before it starts.

Open the Agent page

Help us make Hatch better.

You are one of the first people to use Hatch, and we would like to hear what you think. We read every note.

Hatch is in early testing. Download it, give it to your agent, and tell us what you think. We read every note.

Download Hatch for Mac

Send feedbackThe feedback icon at the top right sends a note with a picture of the window.

You tried 0 of 9 features.