The scenario plays out in B2B software companies everywhere. An enterprise customer wants a specialized dashboard. Another needs a custom automation. A third requests an integration so niche it might serve exactly three users. Product managers face an impossible choice: build it and set a precedent, or decline and watch the customer reconsider their contract at renewal time.
Nour Zahzah heard this story enough times to start looking for an exit. Not from the problem—into it.
His startup, Vendo, emerged from Y Combinator's Summer 2026 batch with an unusual proposition: instead of saying yes or no to these one-off requests, give users the tools to build features themselves. Not in some separate low-code platform bolted onto the side of your product, but embedded directly inside it, using your existing APIs, under your existing permissions. The company released its open-source customization layer this summer.
It's the kind of idea that sounds either obvious or dangerous, depending on who you ask.
Building Inside the Product Itself
The technical mechanism is straightforward enough, at least in concept. Developers install two npm packages and run a handful of commands. Vendo scans your Next.js or Express application, extracting existing server actions and converting them into what the company calls "guarded tools." From there, end-users can request new features using natural language prompts. The system generates UI, assembles micro-apps, and configures automations—all without ever leaving the host product's interface.
These aren't disposable prototypes. Vendo creates structured artifacts it calls "AppDocuments," which can be iteratively refined and, when they prove their worth, promoted to proper server-side code. Users might build a custom dashboard tracking metrics their company uniquely cares about. Or remix existing features in ways the product team never anticipated. Or schedule automations that run overnight, flagging anomalies by morning.
Everything surfaces in an overlay, sidebar, or dedicated panel within the host product's existing design system. For developers who prefer working in their own environment, Vendo can also expose the same tools—and the same guardrails—via the Model Context Protocol, connecting to Claude, ChatGPT, or Cursor.
The pitch is compelling. The implementation is where things get complicated.
The Trust Problem
Zahzah and his co-founder, Yousef Helal, come from different corners of tech. Zahzah studied mechanical engineering at UC Berkeley and played for the San Jose Earthquakes academy before turning fully to startups; Helal holds degrees in electrical engineering and computer science from Berkeley and Stanford, with stints at Apple working on FPGA infrastructure and at AMD on Microblaze V. Before Vendo, they co-founded a startup building AI shopping assistants for physical retail.
That earlier venture put them in constant contact with product teams overwhelmed by feature backlogs—custom requests that would make exactly one customer happy, and thus never quite made it to the top of the roadmap. The pattern repeated too often to ignore.
But letting users generate their own code inside a production SaaS product? That's where most engineering leaders start backing away slowly.
The founders claim Vendo's entire architecture is designed around containment. Every tool receives an explicit risk classification: read, write, or destructive. Destructive operations—anything that could delete data or alter critical settings—always require approval. Automations running on a schedule can't prompt users for permission, so if they need elevated access, they park and wait.
The system employs what Vendo calls "grants," which bind to a specific user, tool hash, operational scope, and time window. If the underlying code changes—say, a developer updates the function—the grant is invalidated. A new approval is required. There's a full audit trail, naturally.
This framework, which the company refers to as "guard.bind," runs through a specific decision sequence designed to catch risky operations before execution. Users can discover additional tools at runtime through a search function, but write operations still demand explicit permission. Everything runs as the signed-in user, under that user's existing permissions.
It's not a free-for-all, in other words. Though whether that's reassuring enough for security-conscious enterprises remains to be seen.
Open Source With a Revenue Model
Vendo has released its core technology under an Apache-2.0 license. The complete agent, an LLM-based policy decision engine, and self-hosted deployment capabilities are available in a public GitHub repository, free forever. No metering, no usage limits in the open-source version. Community support is included.
Revenue comes from Vendo Cloud, which offers collaboration and governance features the free version lacks. The Pro tier runs $49 monthly and includes snapshot sharing, live collaboration, a guard console, and quotas for 2,500 sandbox minutes and 10,000 runs. Teams needing organizational governance, session replay, analytics, and SAML SSO can pay $499 monthly for higher limits: 15,000 sandbox minutes and 100,000 runs. Enterprise deals come with custom pricing, SOC 2 compliance, air-gapped deployment options, SCIM, role-based access control, and dedicated support.
Early adoption numbers are, predictably, modest—the npm package shows downloads in the hundreds, though that figure fluctuates as developers experiment and word spreads through niche communities. GitHub activity suggests incremental refinements and bug fixes in the weeks following the initial release.
Crowded, But Not Saturated
Vendo isn't working in a vacuum. Custom.is markets itself as an AI-powered customization layer for SaaS, enabling business users to build features via natural language prompts. Gigacatalyst, another Y Combinator company from an earlier batch, offers a similar embedded AI layer and claims a 90 percent repeat usage rate among Series B customers—though those numbers are self-reported and lack independent verification.
DronaHQ has been in the embeddable micro-apps space for years, though without the AI-generated UI angle. The broader trend is well-documented: a TechCrunch analysis earlier this year framed the rise of user-built "micro apps" as a fundamental shift in how people interact with software, particularly in enterprise contexts.
Vendo fits that narrative, but with a specific bet. Rather than creating standalone builders or separate platforms, the company is wagering that embedding customization directly into existing products—using familiar interfaces, under existing security models—will feel less jarring to both users and administrators.
Maybe. The challenge is that "letting users write code" still sounds alarming, even with guardrails.
What Happens Next
The team is small—two founders, both based in San Francisco, actively hiring. A careers page invites interested engineers to reach out. Funding details beyond Y Combinator backing haven't been publicly disclosed, and it's unclear what the company's burn rate or runway looks like. An earlier iteration of the startup, which focused on one-click deployment for open-source applications, appears to have been shelved in favor of the current customization platform.
For now, Vendo is making a straightforward bet: that SaaS companies are exhausted by feature request backlogs, and that empowering users to solve their own problems—safely, inside the product—beats both saying yes to everything and saying no to most things.
The open-source release is live. Developers can install it with two commands. Whether it gains traction depends on a more fundamental question: how many product teams are genuinely willing to cede that level of control, even with audit trails and permission gates?
The answer probably says more about the state of B2B software than about Vendo specifically. If the backlog problem really is as universal as Zahzah and Helal believe, then maybe the appetite for solutions—even unconventional ones—is higher than it seems.
Or maybe enterprises will keep saying no, and the backlog will keep growing. Either way, it's an experiment worth watching.
