willmarple/harness-visualizer

This project is a local tool that lets you see, in one screen, what configuration your AI coding assistants are actually reading. Setups for tools like Claude Code, Cursor, Codex CLI, Pi, Windsurf, Copilot, Continue, Cline, Aider, Gemini CLI, and Roo end up spread across many folders: a .claude directory, a .cursor directory, copilot-instructions files, MCP server configs, hook scripts, sub-agent files, and so on. After a few months of adding rules and skills, it gets hard to tell which file outranks which, what is loaded when, and whether two agents are being told contradictory things. Point the app at a project on disk and it inventories every relevant file, grouping results by which agent reads each one. The categories it scans for are context files such as CLAUDE.md or AGENTS.md or cursor rules, skill folders, MCP server configs, hooks defined inside settings.json, and sub-agent files. The same physical file can show up under more than one category. The interface has three tabs. Agents is the default and shows a collapsible panel per detected agent with its config files, scopes, and capabilities. Issues shows aggregated warnings, like an Inspections panel in an IDE. Table is a flat, sortable inventory for power users. Clicking any entry opens an Inspector on the right with a typed Details view and a Raw JSON view. Toggling Edit swaps in a markdown editor that writes back to disk, and a file watcher re-renders the view live. You install it by cloning the repo, running npm install, and starting npm run dev:safe, which checks ports 3001 and 5173 first and asks before killing whatever is squatting on them. The frontend opens at localhost:5173 and the backend stays bound to 127.0.0.1 only, so it is never exposed on the network. There are keyboard shortcuts, atomic file writes, and a six-extension allow-list with a 1 MB cap. The project also ships a portable skill that you install into Claude Code, Cursor, Codex, or Pi. Once installed, the agent reads a generated manifest of the current state and is taught to propose changes through the running visualizer instead of editing the config files directly, so the human can review proposals in the browser. The stack is Vue 3.5, Vite 8, Pinia, Tailwind v4, Express 5, Socket.IO, chokidar, and zod, organized as an npm workspaces monorepo with frontend, backend, and shared packages, plus around 590 Vitest tests.