Account
Explore
In your dashboard and on your phone. Ask once, for one site or all of them: plan, build, review and operate. Respira for WordPress gives AER native builder access, approvals, receipts and undo.
AER keeps the operation in the Respira dashboard. It reads before it writes, waits for your word unless you switch it to Auto, has Jev check every change it makes, and says plainly when something did not work.
Respira AER is included with every paid plan and trial. It works with Respira for WordPress 8.9.0+; 9.0 adds its on-site readiness and entry points.
AER is not a generic chat wrapped around WordPress. Choose a site and Respira for WordPress brings its builder grammar, editable structures, site memory, art direction, available tools, permissions, approvals, snapshots and receipts into the conversation.
Two measured builds of one designed page on mihai.love, a Divi 5 site, not estimates. The first, on 10 September, was Claude Code working through Respira MCP: it set the native Divi target and shaped the fidelity converter. Five days later AER rebuilt the same page from the dashboard in one Auto mode run. Both pages are live below.
15 SEPTEMBER 2026 · CURRENT AER RUN
Nine writes in Auto mode: eight picture uploads and one conversion. AER then read the page back, minted a share link and checked the render through it, all inside the run's four-minute window.
10 SEPTEMBER 2026 · REFERENCE MCP RUN
The first build put the whole page in text modules. The second rebuilt it with native Divi modules, then rendered the page, compared it with the file and corrected the differences.
All four fidelity scores come from the same script. It renders the source file and the built page in Chrome at 1440 and 390 pixels and compares heights, grids, classes, headings and images; the AER and Divi AI scores were measured on 16 September and the converter's on 18 September, against the live pages. The AER receipt stops at the rendered check: its closing message ran past the four-minute worker window after the page was built, and the retry and diagnostic runs that followed are not in the price above.
16 SEPTEMBER 2026 · DIVI AI, ONE PASS
One pass, same site, same file. It writes real Divi blocks and then makes its own page: navy where the design is near-black, Montserrat where it is DM Sans, none of the source's 18 grids, and none of its pictures. It ends up about two thirds of the height.
18 SEPTEMBER 2026 · RESPIRA CONVERTER, ONE CALL · EXPERIMENTAL
The file straight into Respira's converter, the step AER and Claude Code both call. It lays out the same 11 sections, 29 columns and 18 grids, and uploads the eight pictures the file carries. It converts the file it is given and nothing more: choosing what to build and where, checking the render and fixing what is off is the work AER does around it. Using the converter on its own (upload a file, get a native page, no AI involved) is an experimental feature still in the works.
Four steps, the same on every site and every builder.
Pick the sites, type what you want done, on one site or on all of them. Attach the brief, the deck or the zip, and type again while it runs to steer it.
A site can expose hundreds of tools. A small model reads the request and hands the big model only the few it needs.
Page and content edits go to a duplicate first. In Ask mode each write waits for your yes, Plan mode proposes the steps, and Auto mode runs without asking once you have chosen it. The copy's preview carries an approve bar, and an approval link lets a client sign off from the page with no login.
Change runs leave a receipt. Reversible content work keeps the snapshot or undo needed to put it back, and site memory starts the next run further along.
AER is the workspace in your Respira dashboard. Respira for WordPress is installed on each site and gives that workspace secure access to the site's builders, memory, approvals, snapshots and receipts.
Version 9.0 adds readiness checks and an Open in AER entry point to wp-admin. It does not put AER inside wp-admin.
Read the 9.0 release notesRead the AER launch storyRead the AER release notes
Choose the job, add the detail that changes, and review the exact instruction before it reaches the chat. No prompt syntax, no guessing, and nothing runs yet.
In this example you choose Work on one page, then Run an SEO sweep. Each option says whether it reads, writes, or works on a duplicate.
The SEO sweep needs one thing: the page. Enter /pricing. AER asks only for information the finished prompt will actually use.
AER inserts /pricing into the full SEO instruction. Use this prompt places it in the composer, where you can still edit it before sending.
MCP server/pricing. AER wrote the complete instruction below it. “Use this prompt” adds it to the chat; it does not run it.Already know the command? Skip the guide and type it directly. The guided passes and the ready-made prompts come from the same Prompt Book AER uses, so the shortcut and the full instruction stay in sync.
WordPress already has plenty of chats tied to one site. AER is the workspace above the sites: one place for cross-site context, ongoing work, approvals, client feedback, security and review.
The dashboard keeps the threads, sites, plans, previews, receipts and work that continues after the tab closes together. On a phone, it is the same operation at a smaller width, not a reduced companion product.
Claude, Codex, Cursor and other agents can still connect directly to Respira for WordPress when they are the right execution surface. AER remains the home for the WordPress operation; the plugin keeps the same site-side guardrails whichever interface starts the work.
Across your sites
Pick the sites in the scope bar and ask once. Each site gets its own plan and its own receipt.
A fully enabled site exposes 250+ core tools and abilities, and 104 more with WooCommerce. A small model reads each request and passes the big model only the ones that matter for it.
Every run leaves the site memory richer, so the next answer starts from what is already known.
17 builders and WooCommerce, each written in its own format, so pages still open in their editor.
The direction a site is held to goes to the model with the request, so visual work uses your colours, type and spacing instead of inventing some. The active direction also rides in the site context every connected agent reads, so Claude Code sees the same one.
Cards that need no model
Every site's roster of updates in one card. Tick the sites, press Update, and that press is the approval: each site runs its batch and hands back a receipt. No model in the loop, so the rows are exactly what the site said.
The audit's known vulnerabilities that an update closes, on one site or a fleet. Each update is guarded: preflight, a restore point through the site's backup plugin, the update, four health checks, a rollback from wordpress.org when a check fails, and a receipt. A site with no backup plugin says so before anyone presses.
Every Respira copy waiting on one site or a fleet, decided from one card: approve, reject, or send it for approval, with a receipt per site.
Which backup plugin each site runs, its tier and when it last ran. Tick the sites and take a backup now: Duplicator and BackWPup are taken and verified, UpdraftPlus is started and watched, and a site with another plugin or none is named as a finding.
The open review comments on one site or a fleet, grouped by page. Resolve, reply, or apply them on a copy of the page, with a receipt per site.
You stay in control
Ask confirms each write. Plan answers with steps you run one at a time or all at once. Auto runs every write without asking, the site's own confirmations included, for the runs you trust. Plan is the same idea the Respira MCP offers every client: a write comes back as a step with its predicted effect, and nothing lands until the plan is applied.
A proposal becomes a plan card with steps you run, skip, or run all at once.
The preview panel renders the duplicate at desktop and phone width, and says plainly that the live page is untouched.
A receipt names what changed and what it cost. Reversible content work keeps the snapshot or undo that puts it back.
A second model from TypeSafe reads each content change: did it land, is it what you asked, did anything else move. When the answer is no, your model reads the page back before it says done. Included on every plan.
Plain words and no hype. It reads before it writes, tells you what failed and what it tried instead, and never dresses a thin run up as a good one.
Something broke? The R&D tab files a report with the thread's state attached: model, mode, sites and their plugin versions, the last messages and the run log. You see it all before it goes, and it is optional.
Bring the work in
Up to 20 files and 24 MB a message: zip files and folders, docx, doc, odt, pptx, xlsx, rtf, rtfd, pages, md, images and PDFs. Text, links and formatting come out, and any site tool can take an attached file by its name.
Paste or attach a page built anywhere else and it comes back as native blocks the builder can edit, never markup parked in a code widget.
The composer has a microphone. It uses the browser's own speech recognition, the words land in the box as you talk, and nothing is sent until you press send. Respira never receives the audio.
Type / for health, security, accessibility and SEO passes, the six passes judged by Jev, and your own playbooks.
Type while it works. The run reads your note at its next step and changes course without starting over.
Set how hard the model thinks per thread, from a quick pass to a long, careful one.
Bring your own provider key. Respira adds nothing to what your provider bills.
Keep it going
Hand a run off and leave. It finishes without you and emails a report, with a link straight to anything still waiting on your word. A monthly allowance from Builder up.
Turn a run you have watched work into a weekly one. A repeat cannot be armed from a run that has never finished.
Add it to the home screen like an app, with no app store in between, and run the same desk from your pocket, approvals and dictation included.
Type /run review-link and a client opens the page with no login and leaves notes on it, typed or spoken. /run review-feedback turns those notes into a plan, one step per page, applied on duplicates you approve. Builder plan and up.
A run normally lives in the tab you started it in. Turn on Cloud and it stops depending on you: a worker takes custody, and the report arrives by email when it finishes, with a link to approve anything still waiting.
Cloud runs, and runs across several sites at once, come with a monthly allowance: none on Maker, 20 on Builder, 40 on Builder 10 and 150 on Studio, one and a half times that on annual billing, and a trial gets the Builder allowance. It resets on the first of the month, and the app shows what is left. There is no extra charge: the allowance is part of the plan, and the provider bills the tokens.
Bring an Anthropic, OpenAI, Google or xAI API key, or connect OpenRouter in one click. Keys share one sealed vault, you can change the model at any point in a thread and keep the work so far, and recorded costs sit next to your other AI spend in Earn.
Anthropic 
OpenAI 
Google 
OpenRouter 
xAI · Grok 
Jev is TypeSafe's System One decision model. It never writes a word: typed questions go in, probabilities come out, in about half a second. AER asks it what a program cannot decide and a large model should not be paid to decide.
Included on every plan, no extra cost
Every plan with AER has Jev, with no limit for now.
THE WRITE CHECK
After every content write AER makes, three questions go to Jev about the tool's own result.
When one comes back no, AER tells your model to read the page back before it says the work is done. Nothing is blocked, no card is raised, and Auto mode stays automatic. In my own runs a result that read "0 elements updated" scored 0.02 on landing and was flagged; a real change scored 0.96. Each check takes about 0.8 seconds and costs about $0.00004, shown on its own Jev line of the receipt and never added to your model bill.
JUDGED BY JEV · SIX PASSES
A pass gathers a bounded set with the site's own read tools, has Jev answer one question per item, and your model reports what to fix, ranked. Grouped by the job they do.
SEO and content
/run intent Does each page answer what its title was searched for. Worst first. /run promises Does the text under each headline deliver what the headline says. /run links Which pages should link to each other, with anchor words taken from the page itself. Comments and reviews
/run comments Spam, questions, praise, and the comments to reply to today. /run murmur-triage Client review comments by kind and urgency, blockers first. Store
/run catalog WooCommerce descriptions judged for their category and for what a buyer asks before buying. Measured on my own sites, from the receipts: 20 pages judged for search intent in 0.7 seconds for $0.00015, and 20 comments triaged in 0.8 seconds for $0.00016. Find outdated pages (/run stale) uses Jev too when it is on, and your model can call the same judge itself: respira_judge hands Jev up to 200 items and one question.
Both operate the same Respira-connected sites. The choice is where you want the workspace to live, which account pays for the model, and whether Respira's servers should carry the conversation.
| Question | Respira AER | Respira MCP in your AI agent |
|---|---|---|
| Best when | You want one WordPress workspace for many sites, with threads, plans, previews, approvals, receipts, client feedback and cloud hand-off together. | You already work in Claude, ChatGPT, Codex, Cursor, Conductor or another agent and want WordPress to become one of its tools. |
| Respira plan | Included with every paid Respira plan and trial. Cloud and multi-site allowances vary by plan. | Works while the site has an active Respira plan or trial. The Respira site limits of that plan still apply. |
| AI subscription | Not the general launch path. A personal ChatGPT-plan bridge through a local Codex runtime exists as a limited test. | Yes, when the agent includes model use in its plan: for example Claude's own tools on Claude Pro or Max, or Codex on an eligible ChatGPT plan. Client availability and limits belong to that provider. |
| AI API key | Yes for the public launch: Anthropic, OpenAI, Google or xAI, or an OpenRouter connection. It is sealed in your Respira account and opened only inside a run. | Usually no separate AI key when your agent subscription supplies the model. The site connection itself uses Respira browser sign-in or a scoped site access key; that is not an AI billing key. |
| What passes through Respira | Your prompt, selected site context, tool calls and tool results. Threads and run records stay in your Respira account until you delete them or request account deletion. | Not the conversation or WordPress content. Direct clients call the site's own MCP endpoint. Respira still handles account, licensing and, for browser sign-in, authorization metadata. |
| What reaches the AI provider | The prompt and the site context needed for the run, sent by Respira to the provider you selected under that provider's terms. | Whatever your AI agent includes in its conversation and tool context, sent directly under that agent provider's terms. |
| Runs after you close it | Yes on Cloud, within the plan allowance. Approvals can keep waiting for you. | Only if your chosen agent provides its own background or cloud execution. |
Neither path is the hidden “better” one. Use AER when the WordPress operation itself needs a home. Use direct MCP when your existing agent is already the right home. The MCP setup page gives each supported client its actual connection steps.
Elementor, Divi, Bricks, Gutenberg, WPBakery and 12 more. Respira writes in each builder's own format, so the page still opens in its editor.
Gutenberg 
WooCommerce 
Elementor 
Bricks 
Oxygen 
Beaver Builder 
Breakdance 
Divi 
WPBakery 
Visual Composer 
Brizy 
Thrive Architect 
Spectra 
Kadence Blocks 
GenerateBlocks 
GreenShift 
SeedProd Running a store? The WooCommerce add-on brings 104 store tools. See the WooCommerce add-on
Everything here gets written down as it ships, including what did not work.
Respira AER runs on Respira's servers. Your conversations, attachments and the site context returned by tools stay in your account so runs can continue and you can return to them. Delete a thread to remove its stored conversation. Operational access is restricted and logged, and Respira does not use this content to train models. Your provider key is sealed at rest and opened only inside a run. With Jev checks on, your request and each write's result also go to TypeSafe for the check; TypeSafe does not train on them, and the switch is in AER settings.
The plugin and direct MCP path are different: your own agent calls your site's MCP endpoint. Conversation and WordPress content do not pass through Respira, although Respira still handles licensing and may handle browser authorization metadata. In either path, content given to an AI provider is governed by that provider's terms. Read the full data policy.
For the public AER launch, yes: paste a key from Anthropic, OpenAI, Google or xAI, or connect OpenRouter in one click. The provider bills what a run uses, and Respira adds nothing to it. A Claude, ChatGPT or SuperGrok subscription does not include API use. Respira AER seals the key at rest and opens it only inside a run. A personal ChatGPT-plan bridge through a local Codex runtime exists as a limited test, but it is not the general launch path.
Every paid plan and every trial. Attended runs, the ones you watch on one site, are included on all of them. Runs that carry on after the tab closes, and runs across several sites at once, come from a monthly allowance: none on Maker, 20 on Builder, 40 on Builder 10 and 150 on Studio. Annual billing gives one and a half times that, and a trial gets the Builder allowance.
Only after you choose the mode that permits it. In Ask mode page and content edits land on a copy and wait for your yes, and Plan mode only proposes. Auto mode is the one you switch on yourself: it runs without asking. Change runs keep receipts, and reversible content work keeps its snapshot or undo.
Yes, on the Builder plan and up. /run review-link makes a Murmur link: the client opens the page, taps a spot and leaves a note, typed or spoken. /run review-feedback gathers the notes page by page and applies them on duplicates you approve, and the /reviews card lists every open comment on one site or a fleet, page by page, to resolve, reply or apply on a copy.
Five cards: /updates, /security, /approvals, /backups and /reviews. Each asks the sites for their rows, shows them in a card, and the press is the approval: tick, press, receipt per site, on one site or a fleet. A model would only have to copy what the site said, so none is in the loop.
Yes, through the guarded update on the /security card and on the Security page of your dashboard. Before anything is written it checks the offer, that the previous version is archived on wordpress.org, the new version's requirements, governance and the site's backup plugin. Duplicator and BackWPup take a fresh restore point and the run waits for it, UpdraftPlus is started and watched, and a site with no backup plugin is named as a finding. After the update come four health checks, a rollback from wordpress.org if one fails, and a receipt in the activity log.
Respira for WordPress 8.9.0 or later. The plugin is the secure site runtime AER operates through; AER itself lives in the Respira dashboard. The direct agent route runs on MCP server 8.4.19. A fully enabled site exposes 250+ core tools and abilities, and 104 more with WooCommerce.
The current Anthropic, OpenAI, Google, xAI and OpenRouter catalogues, with the models that do the work best listed first. The picker shows only the providers you have connected.
Yes, at any point, even to another provider. The conversation and the work so far carry over: the pages it read, the changes it made, the copies waiting for approval. The next turn runs on the model you picked. Start a job on one model and finish it on another when the first one gets stuck or costs more than the job needs.
A small decision model from TypeSafe. It answers typed questions with a probability instead of writing text, in about half a second. After each content change, AER asks it whether the change landed, whether it matches what you asked for, and whether anything else was touched; when the answer is no, your model is told to read the page back before saying the work is done. Six passes use it to judge a whole site in seconds: search intent, headline promises, internal links, comment triage, Murmur triage and a WooCommerce catalog check, under the Judged by Jev tab on the home screen and at the top of the Do tab. It is included on every plan at no extra cost, every receipt shows the Jev line, and the switch is in AER settings.
Yes. Open it in the browser and add it to the home screen. Same threads, same plans, same approvals, and a microphone in the composer so you can say the request instead of typing it.
Maker is €9 a month after the trial. No card for the trial. AER is included, and it runs on your own provider key: the provider bills the tokens at its price and Respira adds nothing to it.
LLM-ready prompts about Respira MCP + WebMCP
Ask your favorite AI
Opens in a new tab. These prompts are designed to teach LLMs what Respira can do.
