runvendo/vendo

What it solves

Vendo provides a customization layer that allows end-users to build their own features, micro-apps, and views directly within a host product. It solves the problem of rigid software interfaces by enabling users to automate workflows and remix UI components using natural language, all while remaining within the product's brand and security guardrails.

How it works

Vendo integrates a streaming agent into a product using any AI SDK LanguageModel. The process follows three main steps:

1. Extract: It reads the host product's API and converts it into tools that the agent can execute on behalf of the signed-in user.
2. Generate: The agent creates views and user-owned apps from format-tagged UI documents. These components run in a secure iframe jail or a sandboxed server.
3. Guard: A centralized execution choke point manages policies, approvals, grants, and audits to ensure safety.

It uses PGlite (or Postgres in production) for storage and provides a set of modular packages for core logic, agent loops, action execution, and UI hooks.

Who it’s for

Software developers and product teams who want to give their users the ability to customize their own experience, build internal automations, and create custom views without requiring the product team to build every single feature request manually.

Highlights

• User-Driven UI: Users can describe changes to a component or ask for a specific view, and the agent composes it in real-time.
• Natural Language Automation: Users can create standing automations across tools using plain language, with per-tool approval gates.
• Secure Execution: Employs iframe jails and a guarded tool proxy to prevent unauthorized actions.
• Modular Architecture: Provided as a suite of packages (e.g., @vendoai/core, @vendoai/guard, @vendoai/agent) for flexible implementation.
• Voice Integration: Supports live voice sessions where the agent can talk back and render views simultaneously.