NETWORK101

Localmcp — AI skill for Claude Code

AI community

A local, scoped, read-first browser for AI agents (MCP + CLI).

How to install Localmcp

This entry records only its repository, not the path inside it, so there is no exact command to give. Open NETWORK101/localmcp and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Localmcp does

A local, scoped, read-first browser for AI agents (MCP + CLI). The page itself, not a summary.

Alternatives in AI

  • UltraCode Shim — Give Claude Code's ultracode mode to ANY model you already pay for 422 ★
  • Yt Transcriber — CLI app- Give it a YouTube URL and you get a transcription with possible speaker identification and optional s 337 ★
  • Engrim — The Universal Cross-Model Episodic Memory Standard 219 ★

README

localmcp

**A local, read-optimized browser for AI agents.** Your agent reads any page — including the ones you're signed in to — as focused, token-budgeted markdown. Runs a real Chromium on your machine. No cloud relay, no API key, no account.

claude mcp add localmcp -- npx -y @network101/localmcp   # Claude Code
npx @network101/localmcp init                            # everything else: prints config for your clients

[![npm](https://img.shields.io/npm/v/localmcp)](https://www.npmjs.com/package/localmcp) · MIT · Node ≥ 20 · [Website](https://localmcp.pages.dev)


Why

Option Problem for agents
Built-in web fetch (e.g. Claude Code's WebFetch) No cookies, no JavaScript, refuses localhost, and large pages come back as a small-model summary rather than the page.
Cloud scrapers (Firecrawl, Jina, Browserbase) Your Stripe dashboard, internal wiki, and localhost:3000 go through someone else's servers — or can't be reached at all.
Full automation MCPs (e.g. Playwright MCP) Great for driving a browser. For reading, one page snapshot is 13k–87k tokens of accessibility tree (measured).
CLI browser tools Need a shell. Claude Desktop and other sandboxed clients don't have one.

localmcp is the reading-first option: the page itself — no summarizer in the middle — focused and under a token budget, read by a browser on your machine. Five MCP tools, or three shell commands. See the [benchmark](benchmarks/results.md).

What's new in 0.2

  • Real authenticated browsing. npx @network101/localmcp login opens a visible browser on a private profile. Sign in once; the agent reuses the session. Or attach to your own Chrome over CDP.
  • Focus + budget. browse({ url, focus: "rate limits", maxTokens: 1500 }) ranks sections by relevance and returns only what fits — and tells the agent which sections it left out.
  • Structured extraction with MCP sampling. extract returns JSON-LD,