Stripe's Harbor lets an agent write prototypes that each viewer compiles in their own browser
Stripe says more than 3,000 employees have built 12,000 prototypes in Harbor since May 2026. Each prototype is source code that the viewer's own browser compiles on every open, so the build cost lands on viewers' machines.
The Engineer · Build desk
Key figures: 25% of Stripes have created at least one prototype in Harbor since May 2026, per Stripe. 12,000 prototypes were made by 3,000+ creators. One engineer built the first version of Harbor in about 40 days.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Stripes who have created at least one Harbor prototype | 25% | 11 | |
| Made by 3,000+ creators | 12,000 prototypes | 17 | |
| First Harbor, one engineer | 40 days, about | 3 |
What happened
- The stripe.com team often wants evidence on whether a design idea is feasible but cannot always afford to have engineers build a complete prototype.
- Harbor's first version took a single engineer roughly 40 days to build.
- Harbor's agent can read the rendered page, click, type, scroll, switch between mobile, tablet and desktop widths, and take screenshots.
- In one of the first projects, a designer built a product page with an animated hero, and the animation carried over nearly one-for-one when the page shipped.
- After the agent edits a prototype, each comment is reported as attached, needs review or orphaned.
Why it matters
- What it costs cost Viewers' browsers pay for the compile on every open. How that holds up on large prototypes is untested in the public record.
- What it limits constraint With one published version at a time, comparing two design options probably means keeping them as separate prototypes or remixed copies.
- Who's exposed exposure Harbor is described with mock data, yet finance shows monthly actuals in it and a risk analyst built a handle time dashboard. That makes access to a shared URL the open question.
- What it sets up precedent Prototypes are becoming inputs to reviews, analyses and production work, and users are asking for collaboration and links to other AI tools.
Stripe's own wording on the gap is mild. It wrote: "An area where we haven't seen as many improvements from AI has been prototyping." [1] The same post says Stripe built Minions, which write code, and a Knowledge AI Platform for knowledge work, and that many tasks are now faster [16].
Harbor starts from a design system. The author selects one, and that choice sets the packages and starting files before the author describes what they want [5]. From there the agent writes files, Drydock compiles them in the browser, the agent inspects and iterates, and viewers compile the published copy [9]. The hidden dependency is the design system. Stripe's prototypes look like Stripe because the agent starts from Stripe's packages and files and works with realistic mock data [4]. A team without a maintained design system has nothing to select.
Comments are the part of the design we would read first. Anchoring a comment to a page the agent keeps rewriting is, in Stripe's word, challenging [14]. For each comment Harbor keeps several anchors: the element's semantic label and stable ID, its text, its structural path and list keys, and its context, for instance a dialog or accordion that has to be opened before the element is visible [14]. It also records whether each signal was unique when captured [14]. We'd expect that flag to matter most. A text match that was one of several identical buttons is weak evidence after a rewrite, and a unique stable identifier is strong. The resolver then reports a status instead of silently moving the comment [15].
The adoption figures set a ceiling on depth. Stripe reports that 25% of Stripes have made at least one prototype [11]. Twelve thousand prototypes across 3,000-plus creators works out to at most four each [17].
All of these numbers come from Stripe's own blog. Stripe wrote that Harbor "immediately improved the design phase of stripe.com projects, increasing engagement and driving clarity around the final vision" [20]. No metric is attached to that sentence, and the project the post describes in detail is a single stripe.com page [19]. For the numbers to transfer, a company needs a design system an agent can start from and colleagues willing to open a link.
What to watch
- The post's promised section on integrating Harbor with other AI tooling at Stripe, and how a prototype moves between tools.
- Any figure for compile time or source-tree size on large published prototypes.
- Whether the 25% adoption figure, counted since May 2026, holds as the early adopters' novelty fades.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption55
- Hype gap+15
- Incentives60
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Stripe wrote: "An area where we haven't seen as many improvements from AI has been prototyping."
- [2]
When making changes to stripe.com, Stripe's team often wants more information to evaluate the feasibility of a new design idea but cannot always invest in developing a full engineering prototype.
- [3]
One engineer built the first version of Harbor in about 40 days.
- [4]
Harbor is an interactive, browser-based platform where designers use AI to develop shareable prototypes using Stripe's design systems and realistic mock data.
- [5]
The author selects a design system for the prototype, which sets its packages and starting files, then describes what they want.
- [6]
The agent writes a multi-file source tree, and the page compiles and renders in the browser rather than on a server.
- [7]
The agent can read the rendered page, list and inspect elements, click, type, scroll, switch the viewport between mobile, tablet, and desktop widths, and take screenshots.
- [8]
When a prototype is ready the author publishes it to a shareable URL. There is one published version at a time, and every view compiles that copy fresh in the viewer's own browser.
- [9]
The agent writes files; Drydock compiles them in the browser; the agent inspects and iterates; viewers compile the copy.
- [10]
Finance teams started presenting monthly actuals and business reviews as clickable prototypes instead of slide decks, a risk operations analyst turned a handle time analysis into a dashboard the rest of the team could inspect, and account executives built pricing proposals, battlecards and ROI calculators for specific prospects.
- [11]
Since May 2026, 3,000+ Stripes have created 12,000 prototypes, and 25% of Stripes have created at least one prototype in Harbor.
- [12]
As prototypes became inputs to reviews, analyses and production work, Stripes needed better ways to discuss changing designs and move them between AI tools; adoption brought more requests for collaboration within a prototype and interoperability with other AI tooling.
- [13]
Authors can give permission for others to comment on an element of the page or remix the published prototype into a copy of their own.
- [14]
Anchoring a comment to a page that the agent keeps rewriting is challenging. Harbor captures several signals, including semantic labels, stable identifiers, text evidence, structural paths, list keys, and the surrounding context, such as a dialog or accordion, that needs to be opened to reveal the element. It also records whether a signal was unique at the time.
- [15]
When the agent updates the prototype, the resolution flow uses the stored information to report whether a comment's anchor element is attached, needs review, or is orphaned, rather than silently moving the comment.
- [16]
Stripe created Minions that write code and built a Knowledge AI Platform to help Stripes with a wide range of knowledge work; many tasks are now faster and easier thanks to these AI assistants.
- [17]
12,000 prototypes across 3,000+ creators is at most four prototypes per creator on average.
- [18]
Because every view compiles the published copy in the viewer's own browser, the compile work runs on each viewer's machine on every open.
- [19]
In one of the first projects to use Harbor, a designer built a new product page experience, including its animated hero. Engineers, product managers and leaders interacted with it in real time, and when the page shipped the animation carried over nearly one-for-one from the prototype.
- [20]
Stripe wrote that Harbor "immediately improved the design phase of stripe.com projects, increasing engagement and driving clarity around the final vision."
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- shared details
stripe.dev
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- AI-assisted prototypingFollow
- In-browser compilationFollow
- Model Context ProtocolFollow