APIs#MCP

Persona

Persona is Runtypelabs' open-source, framework-agnostic AI chat widget in TypeScript under MIT, 248 GitHub stars, with native WebMCP client tool execution.

Product facts

Commercial Suite
Official sitegithub.com/runtypelabs/persona
Pricing
Free / Open Source
Data checked
2026-10-06

Capabilities

  • Vanilla JS zero-framework dependency
  • WebMCP client tool discovery & execution
  • Streaming replies with voice input/output
  • Plugin system customizes every UI layer

Snapshot figures reflect the check date and may change over time.

Persona is an open-source AI chat UI widget built by Runtypelabs: written in TypeScript, dependency-free on any frontend framework, MIT-licensed, and sitting at 248 GitHub stars as of 2026-10-06 with its most recent commit on October 4. The npm package is @runtypelabs/persona; the maintainers describe it as building “agentic front-end experiences for the web in VanillaJS.”

Core Purpose

Persona solves a specific problem: embedding a customizable AI chat window into any site or product without pulling in React, Vue, or any particular framework. It renders straight to the DOM, so it layers on top of an existing framework app just as easily as it drops into a static HTML page via a single <script> tag. Feature-wise it covers streaming responses, WebMCP page-tool discovery and execution (with a built-in user-confirmation gate, so an agent can’t click a button without approval), built-in local client tools, voice input and output, multi-modal content rendering, tool-call visualization, safe markdown/HTML rendering, and a layered plugin system. It works with any SSE-capable backend, and the repo ships ready-made adapters for Vercel’s AI SDK, the OpenAI Agents SDK, and LangGraph.js, among others. The maintainers are upfront about the boundary: if you’re already building AI interfaces in JSX, they point you to Assistant UI, CopilotKit, or Vercel’s AI Elements instead — Persona is for teams who’d rather not write JSX and just want plain TS/JS hooks. Unlike OpenCrab, which also leans on MCP but focuses on knowledge retrieval, Persona lives entirely at the front-end embedding layer and has nothing to do with document ingestion or knowledge graphs.

Key Scenarios

  1. Static-site intelligence: drop a <script> tag into a docs site or marketing page to add a streaming AI Q&A sidebar.
  2. Cross-framework in-app copilots: layer a chat widget on top of an existing React or Vue product without refactoring it.
  3. WebMCP tool execution: let an AI agent safely trigger actions a page already exposes (cart, filters, form submission), gated by user confirmation.
  4. Custom backend orchestration: the companion @runtypelabs/persona-proxy package, a Hono-based proxy, deploys separately to handle flow configuration for teams that don’t want orchestration logic living in the frontend.

Trying it and its limits

Clone the repo and run pnpm dev from the root; the Vite dev server loads the 35+ interactive demo pages under apps/web immediately, and widget changes apply without publishing to npm. The hosted showcase lives at persona-chat.dev. For a real project, npm install @runtypelabs/persona is enough to get started. Two honest limits: the project is still moving fast (commits as recent as early October), so pin a version before shipping to production, since the API can shift; and WebMCP itself is young — a host page has to implement the WebMCP interface before Persona can discover and run its tools, so not every site you embed it on will expose anything yet. Documentation lives mainly in the repo’s README and examples rather than a dedicated docs site.