Claude Code harness visualizer
A local capability map for Claude Code.
It collects the skills, MCP servers, sub-agents, and plugins you have added to Claude Code and lays them out as “what I can do right now”. A tool for seeing the toolkit you kept growing and stopped being able to track.
$ npx harness-portal
No install. Same command on macOS, Windows, and Linux (WSL) — PowerShell or cmd.
How it works
Three steps. No external server — the only thing that starts is localhost on your own machine.
Reads ~/.claude with an allowlist — only explicitly listed files are opened, and nothing is ever written.
Your own Claude assigns a category to each item. No external server, no API key, no extra billing. Without a Claude Code login it falls back to keyword rules.
Your browser opens on localhost. The first run takes about two minutes for classification and flow generation.
▸ Scanning /Users/you/.claude ▸ Collected skills 48 / MCP 19 / agents 4 / plugins 3 ▸ Classifying 74 items — asking Claude (about a minute) ▸ Classified by Claude ▸ Generating flows … ▸ Generated 10 flows ▸ Saved harness-2026-08-21T02-31-08.417Z.json Harness Portal: http://localhost:4477
What you get
Real screens from a demo harness (48 skills, 19 MCP servers, 4 agents, 3 plugins), run with --lang en. The app also has a JA / EN switch in the header.
Every item, grouped by category. Names are never truncated. An MCP server registered in several projects is shown once with a count badge. Search across everything.
“For this job, these steps” — assembled from the tools you actually have. If no tool covers a step it says “No tool available” instead of pretending. Each flow exports a skill prompt (Markdown) you can hand straight to Claude Code.
Detects MCP servers that carry the same name but different settings across projects, or identical entries registered more than once, and asks “Is that intentional?” — the kind of thing you never notice by scrolling a list.
Private by design
~/.harness/snapshots/. No account, no sign-up.