Persistent Web Skills Runtime

A WebMCP-first runtime that learns browser workflows once, reuses them reliably, and repairs them when the web changes.

Comment

Persistent Web Skills Runtime

Persistent Web Skills Runtime (PWS) is a WebMCP-first Chrome runtime that lets agents learn a browser workflow once, reuse it as a versioned skill, and repair it when the web changes.

The problem

Agentic browser tasks are commonly rediscovered through fragile clicks on every run. Small UI changes can break a workflow, while useful structured WebMCP tools may be ignored or forgotten. Repeated exploration is slow, brittle, token-intensive, and hard to trust.

The solution

PWS routes each task through a clear priority ladder:

Native WebMCP → Cached Web Skill → Browser/DOM learning fallback

It discovers the page's current WebMCP tools first. If there is no suitable native tool, it learns a successful semantic browser workflow, stores it in IndexedDB with immutable version history, validates the outcome on reuse, and repairs only the failed step when a dependency changes.

The live taskboard demo shows an 11-action Learn → Reuse → Repair flow: it learns a create-task workflow, reuses it, survives a Create Task → Add Task UI change with a validated v1 → v2 skill repair, and demonstrates live native tool registration through toolchange.

Key features

  • Native-first WebMCP routing via document.modelContext.getTools(), live toolchange synchronization, and current-tool rebinding.
  • Safe capability snapshots that persist metadata, not executable tool handlers.
  • Semantic locators based on role, name, labels, stable attributes, and meaningful text instead of brittle structural selectors.
  • Dependency-only DOM fingerprints, explicit result validation, and targeted step repair.
  • An inspectable Runtime Inspector, execution log, timings, validation status, and skill versions.

Built with

React, TypeScript, Vite, a Chrome Manifest V3 extension, IndexedDB, Node.js, and an optional OpenAI Responses API service. API keys remain server-side; the interactive demo runs without a key through deterministic local learning and repair adapters.

Career value

I work as an agent engineer and repeatedly test browser automation over many fresh pages. I wanted a way for agents to achieve stronger results with fewer tokens. Building PWS made the practical value of WebMCP clear: expose structured tools, preserve proven workflows as reusable skills, and give agents a safe answer to reuse instead of forcing them to rediscover the UI. This can make agent browser work faster, more reliable, and easier to trust.

Built With

Share this project:

Updates

Submission history