Vendo launches open-source embedded AI agents for end users to build SaaS features
- Vendo is a public GitHub project that lets a product's end users create features and dashboards inside that product with AI, rather than requiring the vendor to build every requested workflow.
- Vendo says generated components run in a sandbox, and its guardrails require displayed data to come from a real API call so the model does not invent product data.
- The project says it learns a host product's design system and brand and uses a Vendo component library to make generated UI look native to the host application.
- Vendo's current pricing includes a $49 monthly plan with $49 in usage credits shared across AI inference and sandbox minutes; the team says its LLM tool-call safety judge is experimental and has no customer deployments.
- The repository lists 503 stars, 74 forks, 1,242 branches, and 3,863 commits at the time captured in the article.
Hacker News opinions
An API plus MCP access already lets my users point Claude at the product, and some have built local UIs with Claude Design. They can use an AI subscription they already pay for, as long as they stay within the MCP's permissions.
I agree there is limited benefit for people who already use Claude Code. We are targeting vertical SaaS companies with many nontechnical users who need Claude Code-like capability inside the product.
I think businesses will need this. Our nontechnical users fork our open source Rust desktop app and use Claude Code to build workflows, while we keep ownership of the hard distributed-systems and performance work.
How far can each dashboard be customized? I want to know how detailed the generated result can get.
Users can get as detailed as they want. We focus on sandboxing, correct data representation, and making generated work feel native rather than restricting them to preset components.
This sounds brutal for support. Customers will create business-critical workarounds, break them, and expect support staff to understand and repair AI-generated features.
We have spent time on that problem, and expect agents to resolve many issues without human intervention, though not all of them. Most users are making noncritical additions today, and we think the customer customization benefit outweighs that risk.
I think software needs to become more malleable for users. Vendors should provide the hard widgets and default UX, then let customers request features through a chatbot and possibly publish popular work in a marketplace.
How do you keep generated UIs good enough to put in front of tech-literate customers who will spot errors?
We require displayed data to come from real API calls, learn the product's design system and brand, and use components built for Vendo's generation flow.
The pricing page is unclear: paying $49 per month for $49 of usage does not tell me what a few hundred sessions means. I am also wary of an LLM safety judge deciding whether tool calls run, ask, or block.
The usage credit covers AI inference and sandbox minutes, and we will make that clearer on the pricing page. The LLM safety judge is still experimental and no customers use it.
I think the durable interface is an API with CLI and MCP access plus a SKILL.md file, such as /api/openapi.json and /api/skill.md. AI can act as the integration layer, and I have seen similar approaches work.
How does the VM handle local state and performance when users fetch, join, and render data from multiple endpoints? In our Dart and Go widget system, output speed and quality mostly depend on the selected inference endpoint.