Now it makes
the images too.

1. Images, both directions

Generate it from what's on the page. Then use it.

The agent now has an image model in its hands. It can create a graphic from the context it just gathered, edit one you already have, and carry the result straight into the next step — because generated images are real files in the run, not previews in a chat window.

That closes a loop that used to need you in the middle: research a topic, make the visual, attach it to the LinkedIn post, publish. One prompt, no download-and-reupload.

And it reads pixels now, on purpose. Take a full-page screenshot and ask a specific question about it — “what price do the three map markers show?” — and you get an answer, not a generic caption. Charts, maps, canvas apps, and rendered layouts stopped being blind spots, even on models without vision.

2. Cloud, rebuilt — and about 2× cheaper

Run several at once. Pay roughly half.

We rebuilt how the agent assembles its context, so most of every step is now served from cache instead of paid for again. Page snapshots shrank 43%, repeated steps reuse what the model already saw, and the agent stopped burning turns re-reading pages it could already see. The result across a typical task: roughly half the cost of last month.

The cloud dashboard got the same treatment: start a fresh chat in one click, run several tasks in parallel without them tripping over each other, and read results as clean markdown, tables, and sheet previews. Sign into a site once by taking over the live browser, and your next cloud run starts already logged in.

3. On AppSumo

Past $10k in lifetime deals.

The AppSumo launch crossed $10k in sales, and the lifetime tiers are still live — monthly credits plus Rover website allowances, activated by signing in with your license. Renewals now roll unused allowance forward correctly, so nothing you paid for evaporates at the end of the month.

4. Meanwhile, Rover

Your forward-deployed engineer, on every visit.

Rover stopped waiting to be asked. It greets visitors with a spoken, personalized welcome, remembers people who've been back before, and offers one-tap live demos you can share as a plain link — the job a forward-deployed engineer does on a first call, running on your site at 3am.

The voice got better too: more natural narration that no longer cuts out mid-sentence, and demos that keep talking even when a click opens a new page or tab.

5. Under the hood

What the agent actually sees.

Retriever doesn't drive from screenshots. Every observation is a compact semantic tree of the page where each element is addressable — which is why a single Amazon results page can hold 8,699 nodes and 1,137 link targets, all of them in reach, when the viewport shows a few percent of that.

We published a deep dive on how those trajectories are captured, and a roadmap for what we do with them next: turning every completed task, every failure, and every human rescue into tested capability your agents keep.

Backed by

NVIDIA Inception  ·  Google Cloud for Startups  ·  Salesforce Launchpad
NEC X Ignite  ·  Bright Data  ·  ElevenLabs Grants  ·  Composio  ·  GMI Cloud SCALE

Thanks to the programs backing this work — and to you for running the agent that keeps teaching it.

Ask for the post and the picture.

Research it, write it, generate the image, publish it — one prompt, in the browser you're already signed into or a cloud one that stays signed in for you.

Happy shipping,
Team rtrvr.ai