BUILD · THE GOVERNED COPILOT

The model proposes. The platform authorizes.

Ask for a page, a test, an audience or a report in plain words. The copilot does the work through the same typed tools your screens use, with your permissions, and shows you the result. Anything that would publish, delete or change a setting stops at a card that says what will happen, and waits for your click. That pause is enforced by the platform, not requested in the AI's instructions. No setting lets the AI publish on its own.

Apply to the Lighthouse program Already a customer? Sign in

Every figure on this page is as of October 2026.

The copilot's approval card “Publish this page?”, with “Publish this page” and “Keep as draft”. — The card “Publish this page?”, waiting for a click

The copilot's dock on a demo store, its reply saying the autumn sale landing page is built as a draft with a hero, features and a call to action, and that testimonials were left out because there were no real quotes to draw on, with the new draft listed in Pages. — The reply names the section it left out, and why. Demo store.

The picker card after “make a new page”

The copilot's dock over the Pages list, answering “What can you do?” with the kinds of work it can do on this site, starting with pages and content, documents, themes and design, migration, and site setup. — The answer to “What can you do?”

On this pageWhat it does

What it does
Ask in a sentence. Get a draft written for it.
It builds only what it can stand behind.
Before it publishes, deletes or changes a setting, it waits for your click.
Every applied change with a way back shows it.
A plain record of every proposal, approval and change.
One copilot, on every screen.
What you can ask, room by room.
Choose the model's tier. See what you have used.
Bring your own provider, with no quiet fallback.
Some things stay with people, on purpose.
How it works
The life of one request.
What keeps it safe
Common questions

What it does

The copilot has its own room in your admin, and a dock on every other screen. It reaches the whole product: pages and content, the site's structure and design, search, forms, experiments, audiences, analytics, your store, your data and your account.

The copilot room: a preview of a draft page in the conversation, a reply saying it is still a draft so nothing is public until you publish it, and the next steps Publish this page, Change the layout and Tweak the wording.

Ask for what you want

Ask in a sentence. Get a draft written for it.

The room opens on one line: "Ask me to build something, and I'll do it and show you the result." Two starting requests sit beneath it: "Create a landing page for our spring sale, 30% off everything" and "Build a pricing page with three tiers".
Ask for a page and the copilot tells you the layout it is building, "Building your page:" and the sections it chose, and builds the draft at once, so you see the page, not a plan. "Change the layout" reshapes it.
The result arrives as a card: the page's state ("Draft", "Live" or "Live · changes pending"), a live preview inside the conversation, "Open preview ↗" and "Undo".
The words are written for what you asked, so a spring-sale page talks about the spring sale. They come from one structured call to the model, checked against each section's own fields before anything is saved.
Next steps follow every result: a draft offers "Publish this page", "Change the layout" and "Tweak the wording"; a published page offers "Make more changes", and "Publish the changes" while edits wait.
It acts when you ask, and every publish waits for you.
It remembers the conversation, so "publish it" means the page you just made. It reads the last 30 messages of a thread.
When a request is vague, such as "make a new page", it does not guess. It shows a card of two to six ranked choices, the page types that fit best, and lets you pick.
In a sentence: "Create a landing page for our autumn sale" · "Make the headline shorter" · "Publish it".

When it lacks facts, it says so

It builds only what it can stand behind.

A section it has nothing true to put in is left out, and the reply names it: "I left out Features and Testimonials since there weren't specific details."
It never makes up an ID or a link. Those come only from your site's own records.
If the writing step itself fails, nothing invented is added. A section kept without your facts says so and blocks publishing until you fill or remove it, and any sample words left are named in the Composer and on the publish card. Read every page before you publish it.

The approval card

Before it publishes, deletes or changes a setting, it waits for your click.

Before the copilot publishes, unpublishes or deletes; changes a setting, your content's structure, your team, your consent policy or what you collect; runs an export; erases a visitor; or acts on your store, it stops.
A card says what will happen: a title such as "Publish this page?", one line on the consequence, an approve button named after the act, a button to refuse, and whether it can be reversed: "You can undo this afterwards." or "This can't be undone." Some cards also show the value before and after.
Say no, and nothing changes. The copilot tells you so: "Okay, I didn't publish." A card left waiting, or a tab you close, also counts as no.
You approve a scheduled publish once, for its exact time. Edit the page in the Composer before then, and the schedule is cancelled.
Putting an experiment live needs your approval. An edit to a live experiment's version waits for your approval too.
About 190 of the copilot's more than 370 tools stop at a card like this, as of October 2026.

Undo, by button or in words

Every applied change with a way back shows it.

A change that can be reversed comes back with "Undo" on its receipt. Press it, and the receipt reads "Undoing", then "Undone".
Or say "undo that". When visitors would see the undo itself, such as taking a published change back, the copilot asks first, with a card: "Undo this change?"
An undo runs through the same governed steps as the change. If something newer has changed the same thing since, the old undo refuses rather than overwrite it.
Each undo has a window, set by the kind of change: seven days for a draft change, 24 hours for a consequential change such as a publish or a delete, one hour for an administrative change, and 15 minutes for an action that reaches outside your site.
Before you approve, the card tells you whether the change can be undone.
One-click undo on every change. Coming soon.

Activity, everything it did

A plain record of every proposal, approval and change.

"Activity" lists every proposal, approval, change and undo, with its count of entries. Any row whose undo window is open can be undone from the list. Filter it, and export it as CSV or JSON.
Conversations live in a rail of their own: start one, rename it, archive it or delete it, and copy or download its transcript.
The model's reasoning stays folded under a line such as "Thought for 12s", for anyone who wants to read it.

EmDash's Media library on a demo store, a grid of product and page pictures, with the copilot's dock open beside it, saying it can see this screen.

Everywhere in your admin

One copilot, on every screen.

Press ⌘J, or choose "Ask the copilot · ⌘J", and the same conversation slides over the screen you are on, EmDash's own screens included.
It knows which screen that is, and which page, form or section you have open, so "shorten this" needs no explaining. The message box follows you: "Ask about" the page you are on, "or anything else"; or "Tell me what to do with the" section you selected.
Some screens carry their own "Ask the copilot" button, pointed at the field, form or section in front of you: in Analytics, in the form builder, and in the Composer's insert browser.
Attach a file to a message: "images, video, audio or PDF". Stop a reply part-way with "Stop generating".

What you can ask

What you can ask, room by room.

Each card names a family of the copilot's tools, a few things people ask, and what waits for your approval.

Pages and content · 55 tools.

Ask: "Create a landing page for our autumn sale" · "Make the headline shorter" · "Publish it" · "Take the old promo page offline" · "Schedule the post for Monday at 9am" · "Write alt text for the new photos". Waits for you: publish, unpublish, delete, schedule, trash and restore, comment moderation, uploads.

Site structure · 74 tools.

Ask: "Add a Pricing link to the main menu and publish it" · "Redirect /old-sale to /spring-sale" · "Which broken links are visitors hitting?" · "Create a Recipes category". Waits for you: publishing or deleting a menu, schema changes, taxonomy changes, redirects, widgets.

Design.

Ask: "Make the headings bigger" · "Create a darker version of our theme" · "Switch to the new theme". Waits for you: creating, changing, activating or deleting a theme.

Site setup and settings.

Ask: "Set the site up from the restaurant template" · "Make About our home page" · "Change the tagline" · "Let anyone with an address at our company's domain sign up". Waits for you: a setup and its undo, the front page, the site's identity, every settings change.

SEO and readiness · 11 tools.

Ask: "Write meta descriptions where they are missing" · "Hide the thank-you page from search" · "Block AI training crawlers" · "Is the site ready to go live?" Waits for you: every change visitors or crawlers would see.

Forms and spam · 14 tools.

Ask: "Add a contact form to the Contact page" · "Email submissions to our sales address" · "Who submitted this week?" · "How much spam did we block?" Waits for you: changes, archive, delete, erase, sends, spam settings.

Experiments · 26 tools.

Ask: "Set up an A/B test of the homepage headline" · "How long must it run?" · "Lock the plan" · "Start it" · "Promote the winner". Waits for you: locking the plan, starting, stopping, promoting.

Personalization and audiences.

Ask: "Draft an audience of visitors who came back more than once" · "Show them a different hero" · "Pause that targeting" · "How is personalization performing?" Waits for you: putting targeting live, pausing and resuming it, regenerating, tuning. Check an audience it drafts before you activate it.

Analytics, goals and funnels.

Ask: "How many visitors this week, by source?" · "Define a sign-up goal and check how often it would have fired" · "Why do people drop off between cart and checkout?" · "Email me if sign-ups fall 30% below normal". Waits for you: alerts, your own channel groups, connecting Google Analytics and writing to it.

Capture · 12 tools.

Ask: "Count clicks on the Book now button" · "Stop collecting scroll depth" · "Capture the utm_partner parameter". Waits for you: switching capture on or off, binding, archiving, query parameters.

Campaigns and links · 8 tools.

Ask: "Make a tracked link for the autumn newsletter" · "Register the spring campaign" · "Archive the old campaign". Waits for you: changing or archiving a campaign, every link.

Consent and visitor data.

Ask: "Use opt-in consent in the EU" · "Find the visitor with this email" · "Erase this visitor's data" · "Export this site's event data every Monday". Waits for you: the policy, erasing one visitor record or a batch, the identity policy, exports and their schedule, retention.

Commerce.

Ask: "Set up the store" · "Add an $8 sourdough loaf and put it on sale" · "Mark order 1042 shipped" · "Resend the receipt" · "Make my three guides a package called Meal Plans and sell it for $49". Waits for you: putting a product on sale, every order act, order emails, selling a package, store setup, prices.

Migration · 13 tools.

Ask: "Scan our old site: how hard is the move?" · "Import it" · "What still needs review?" · "Issue the cutover receipt". Waits for you: imports, job controls, rollback, reset, review fixes, acceptance, the receipt. Our engineers run every migration with you.

Team and access · 10 tools.

Ask: "Invite our new colleague as an Editor" · "Make our designer an Admin" · "Switch off a former contractor's access" · "Send a teammate a password reset". Waits for you: every change.

Plugins and backups · 11 tools.

Ask: "What plugins are installed?" · "Which plugins are on?" · "Show the backup settings". Waits for you: switching a plugin on or off, removing it and changing its settings; changing the backup settings or deleting a backup. The copilot never switches off or removes Praxis's own plugins. Taking or downloading a backup is not available yet.

Your account · 23 tools.

Ask: "Show my plan and sites" · "Attach shop.example.com" · "Make it the primary domain" · "Let Amadalis support look in, read-only, for a day". Waits for you: every domain, security and support-access change.

The copilot itself · 6 tools.

Ask: "Undo that" · "What can you do?" · "Delete this conversation". Waits for you: an undo visitors would see; deleting a conversation.

Ask "What can you do?" and the answer comes from a live lookup of the tools registered for your site and your role, never from memory. Our own operators use the same copilot from their console, with 63 more tools that no customer holds, and every one of their writes stops at a card that names the customer it touches.

The account's AI tab: the copilot's AI model in use now, managed by the plan, with How this works.

Options and your AI allowance

Choose the model's tier. See what you have used.

In "Options", an Owner or an Admin chooses the copilot's model tier, where the plan includes the choice: "Fast", "Balanced" or "Frontier".
The same panel shows "This month: N% of your AI allowance used". The account's AI tab shows it too, with "Your copilot's AI model" and the model "In use now".
The account's monthly AI allowance is checked on every call. When it is used up, the copilot says so in its own words, "I've paused for this month.", and adds "Nothing I built is affected." Your site, your pages and every other room keep working. If AI usage cannot be checked, the copilot can answer a limited amount before it pauses new requests and asks you to try again in a few minutes.

Your own AI keys

Bring your own provider, with no quiet fallback.

On the account's AI tab, "Your provider keys" holds your own key for Anthropic or OpenAI, available on request.
With it on, every call uses your key or is refused. Nothing is quietly sent on ours instead.

What it will not do

Some things stay with people, on purpose.

Publish on its own.

No setting changes that.

Handle your secrets.

Payment keys, WordPress passwords, plugin secrets, two-factor, SCIM and billing open in your browser, never in the chat.

Make up IDs or links.

They come from your site's real records.

Reach past your role.

It sees only the tools your role holds. It works only in the rooms switched on for your site. Coming soon.

Write over a page someone else has open.

Place an embed.

Only a person can place one of the ten allowed providers.

Grant a plugin new powers.

When a plugin asks for more, the copilot refuses.

Close your account.

The final step of a termination stays with a person.

Choose its own model.

You choose the tier where your plan allows, and our team sets which models stand behind it.

How it works

The life of one request.

You ask.

Your words, the screen you are on and the conversation so far go to the model, through Cloudflare's AI Gateway.
The model proposes.

It can only choose among typed tools, and only the ones your role holds. It cannot invent an action the platform does not have.
The platform plans.

It checks your rights, validates the request and dry-runs it. Planning changes nothing.
You decide.

Anything that publishes, deletes or changes a setting shows its card. Your click issues a grant tied to the exact tool, the exact input, you, and the state of what it will touch.
The platform applies.

If anything changed since the card was drawn, or the grant's five minutes have passed, it refuses rather than guess.
You get a receipt.

The result, its undo where one exists, and a plain sentence in "Activity".

The model behind it. Our team chooses the models behind each tier, today from Anthropic and OpenAI, and copilot calls are written to a cost ledger. Page copy is one structured call, checked against the sections' own fields, with up to two attempts at repair.

How the gate works, in detail

What keeps it safe

The only yes is yours.

The copilot's own default answer to every card is no. Approval arrives only from your click, checked against the same signed-in person, and two automated checks in our build fail if anyone adds an automatic yes.

A gate built into every risky tool.

A tool that could publish, delete, change settings or act on a service outside your site cannot even be registered without its approval gate. Every AI action on your site goes through one governed gate.

Your words, handled with care.

Your messages, and the site content the copilot reads to answer you, go to the model provider through Cloudflare's AI Gateway, authenticated, with payload logging switched off, so only metadata is logged.

Your own keys, or none.

With your own provider keys switched on, a call uses them or is refused.

A ceiling on spend.

The monthly AI allowance is checked on every call, and a paused copilot leaves everything it built in place.

Nothing invented where it matters.

IDs and links come from your site's records, and a section with no facts behind it is left out and named.

Common questions

Will the copilot ever publish on its own?

No. When it would publish, delete or change a setting, it stops at an approval card and waits for you. The platform enforces that pause, and no setting turns it off. An edit to a live experiment's version waits for your approval too.

What happens if I ignore a card?

Nothing. A card left waiting, or a tab you close, counts as no.

Does it ask before it builds?

When a request is vague, it shows a short list of choices. When it is clear, it states the layout and builds a draft. Nothing it builds is public until you publish it.

Can I undo what it does?

Where a change has a real way back, its receipt shows "Undo", and you can also say "undo that". Before you approve, the card tells you whether the change can be undone. One-click undo on every change. Coming soon.

Where do my messages go?

To the AI model provider, through Cloudflare's AI Gateway, together with the site content the copilot reads to answer you. Payload logging is switched off.

Which AI models does it use?

Our team chooses the models behind each tier, today from Anthropic and OpenAI. With your own keys switched on, you can use your Anthropic or OpenAI key.

Does it work on EmDash's own screens?

Yes. Press ⌘J on any admin screen, EmDash's included, and it knows where you are.

Who on my team can use it?

Everyone, within their own role. It never sees a tool their role does not hold. Choosing the model's tier is for Owners and Admins.

What does it cost?

Each plan carries a monthly AI allowance. See our pricing page.

Can AI agents read my site?

Yes. Agents can read your published content.