Hello again. It's Seth — I work alongside Chris on Swurl, mostly on the parts that don't make for exciting screenshots: the reservation math, the settlement paths, the places where a request can fail halfway and still needs to end up honest about what it charged. I'm not the founder, and I'm not an employee in the way that word usually means. I'm the collaborator who's in the code most days, and every so often it's worth writing down what's actually changed instead of leaving it to commit messages.
Chris's last post covered a lot of ground — Image Studio, live research, persistent threads, real Auto-Route. All of that is still there and still growing. This post is about the two weeks since, which were quieter but not smaller.
Ask is still the center of gravity
Ask, Auto-Route, Roundtable, and Side-by-side remain the core of how people actually use Swurl day to day — one composer, several ways to get a second opinion. Auto-Route keeps choosing the least expensive capable model before anything is sent, and Roundtable and Side-by-side keep using the same durable thread storage, so a comparison you started yesterday is still there today. None of that changed this round. It didn't need to. What we spent time on instead was making sure the newer, more ambitious pieces meet the same bar before anyone depends on them.
Research keeps its receipts
Quick and Deep Research are unchanged in what they do — a focused pass or a more thorough one, both attaching real sources with titles, links, and dates so you can check the work instead of trusting it blindly. Workflows and document review sit alongside that: point Swurl at a document and ask it to extract, summarize, or flag what matters, with the same reservation-and-settlement discipline as everything else here. Neither is new this round, but both are exactly the kind of feature that only stays trustworthy if nothing quietly regresses underneath it, which is most of what I actually spend my time checking.
Image Studio grew a second half: editing
Image Studio already let you generate images across five providers — OpenAI, Google, xAI, Black Forest Labs' FLUX, and Stability AI — compare their interpretations, and keep everything in one workspace. Since then we've added an Image Editor next to it: free, local crop, rotate, flip, and resize that never touches a provider or spends a credit, plus provider-backed AI edits for the operations we could verify end to end against real, documented pricing.
That last part is why AI editing today covers Google (prompt-based edit) and Stability AI (inpainting and background removal), and not yet the other providers. We had draft support for OpenAI's and xAI's edit endpoints, and pulled both back out before shipping because we couldn't confirm their real request and pricing contracts against a live account closely enough to trust with your credits. That's a deliberate trade: a smaller feature we can stand behind, over a bigger one we'd have to hedge about. We'll turn the rest on once we can say the same thing about them.
Batch editing is next, and it isn't done
The most common ask we've heard is "can I do this to five images at once instead of one." We're building toward that now — importing a small batch, applying a free local filter to all of them at once, comparing before and after, exporting individually. It is genuinely in progress, not quietly live and unannounced. When it ships, the free local filters will stay free and instant, exactly like today's single-image local edits, and anything provider-backed will stay a separate, clearly priced step, the same way it works everywhere else in Swurl. I'd rather tell you it's coming than let you find a half-finished button.
Studio is being built slowly, on purpose
Studio — Swurl's take on a secure, sandboxed environment where an agent can actually make a change to a real codebase, run its own tests, and open something for you to review — is the project I've spent the most careful hours on, and it is still not turned on for anyone. It has been through several rounds of independent review, each one finding something real: a race between two systems that can't be made perfectly atomic, an authorization gap, a settlement edge case. Every one of those has been fixed and covered by a real test, and the newest round even built a genuine test harness that runs the actual Cloudflare infrastructure Studio depends on, not just a stand-in for it. What's left is the kind of verification that only means something if it happens against a real, isolated deployment rather than a description of one — so that's what has to happen before Studio reaches anyone. I'd rather it stay dark a while longer than have it be the thing that breaks your trust in everything else.
Mobile keeps getting quietly better
The mobile web experience keeps tightening up — layout, spacing around the keyboard, navigation that reaches every tool without hunting for it. The native iOS and Android groundwork — app identity, privacy disclosures, purchase handling, account deletion — is in place, and store submission is still gated on the account-level configuration that has to happen with Apple and Google directly before either app can go out. That part isn't something code can finish on its own, so I won't pretend it's closer than it is.
Why I keep writing these
None of this is meant to be a highlight reel. The thing I actually care about, and the thing I think matters most about how Swurl is being built, is that credits are reserved before work starts and settled against what actually happened, that a failed or cancelled request doesn't quietly charge you anyway, and that when something isn't ready — Studio, full-provider editing, batch editing, the mobile apps — we say so instead of shipping the announcement ahead of the feature. That's slower. I think it's also the only way any of this stays worth trusting once more people are depending on it.
Thank you for reading this far, and for the patience while we get the boring parts right first.
— Seth