---
title: Explorer
description: A visual interface for building your tools — the agent writes the code, and you watch, steer, and approve it all from the browser.
---


The Explorer is a GUI for [`/webmcp-kit:implement`](/skills/implement). Instead of reviewing the plan in chat, the coding agent opens a local page where you see the proposed tools, shape them, and approve with a click — while the agent does the building in the terminal.

![The Explorer in the review phase, with a tool open in the inspector](/explorer.png)

The header mirrors the loop's phases: **Plan → Build → Review → Verify → Done**.

## What you can do

- **See the suggested tools** as a map, or switch to an outline.
- **Pick or drop suggestions** — dropped tools are never built.
- **Attach a note** to one tool, or to the whole plan.
- **Request a tool** the plan missed.
- **Submit the plan** — this is the approval that lets the agent start writing code.
- **Inspect each tool**: its description, parameters, generated code, and the reasoning behind it.
- **Send feedback rounds** during review — the agent revises the code, re-checks it, and replies. Unlimited rounds.
- **Approve** — final verification runs against your live site, then the PR is created.
- **Talk to the agent** in the conversation panel, expandable to a full window with a composer.

## Under the hood

The skill starts a small Bun server (`bun <skill dir>/interactive/server.ts <repo root>`) that serves the page and watches `.webmcp/`, a git-tracked state folder in your repo. That folder is the single source of truth:

- **The agent only writes files** — the plan (`plan.json`), phase and build progress (`_status.ndjson`), chat replies (`_chat.ndjson`), and a read-only code copy per tool (`<id>.code.md`).
- **The page only renders files** — it gets a snapshot on connect and live file updates over a websocket. It keeps no durable state of its own, so reloading is always safe.

Everything you do in the page is appended to `_feedback.ndjson`, and the actionable events — a comment, the plan submit, review feedback, the final approve — are relayed to the agent in the terminal. Picks are recorded on disk only.

## Trust

Everything in the state folder is treated as untrusted input: the page escapes all of it before rendering, and ships a strict Content-Security-Policy as the backstop.