tiny-browse

Give any AI agent the web, in 126 kB.

tiny-browse searches Google in the Chrome you already have, reads the best results and returns clean Markdown that fits a hard token budget. One zero-dependency package: an MCP server for any agent and a Node.js library for your own code. No bundled browser, no API keys.

Related MCP server: Neurobird Search MCP server

• Any agent, any model. MCP over stdio or HTTP, plus a JSON API. No LLM inside.

• Tiny on disk. A 126 kB download with zero runtime dependencies, using your installed Chrome, Edge, Brave or Chromium instead of downloading a browser of its own.

• Answers, not HTML. Main content only, with headings, code, tables and lists kept, inside a hard max_tokens limit.

• Careful by default. Strict search rate limits, no access to private networks, pages read without your cookies.

Install

Requires Node.js 22+ and Chrome, Chromium, Edge or Brave (or use the no-Chrome mode).

From npm

npm install -g @shubhamcodess/tiny-browse
tiny-browse doctor

Agents can also run it without installing: npx -y @shubhamcodess/tiny-browse mcp.

From the latest GitHub release

npm install -g https://github.com/shubhamcodess/tiny-browse/releases/latest/download/tiny-browse.tgz

From source

git clone https://github.com/shubhamcodess/tiny-browse.git
cd tiny-browse
npm ci && npm run build
npm link            # or run: node dist/cli.js

Use cases

Web search for your own local agents. An agent running on a local model has no web access. Add tiny-browse as an MCP server and it can search and read pages, with no API keys and no cloud service.

Answers grounded in sources. Every page comes back with its URL and only its relevant sections, so an assistant can answer from current pages and cite them instead of relying on what the model remembers.

Current docs for coding agents. Libraries change after a model's training cutoff. web_fetch with focus pulls just the part of the documentation the agent needs, as clean Markdown with code blocks intact.

Connect to an agent

Claude Code

claude mcp add tiny-browse -- npx -y @shubhamcodess/tiny-browse mcp

VS Code (GitHub Copilot): .vscode/mcp.json

{
  "servers": {
    "tiny-browse": { "type": "stdio", "command": "npx", "args": ["-y", "@shubhamcodess/tiny-browse", "mcp"] }
  }
}

Cursor, Claude Desktop and most other clients

{
  "mcpServers": {
    "tiny-browse": { "command": "npx", "args": ["-y", "@shubhamcodess/tiny-browse", "mcp"] }
  }
}

tiny-browse config <client> prints the setup for claude-code, claude-desktop, cursor, vscode, codex or generic. To share one warm Chrome between several agents, run tiny-browse serve --http 7777 and connect to http://127.0.0.1:7777/mcp.

The agent gets two tools: web_search (search and read the top results) and web_fetch (read URLs, optionally narrowed with focus or continued with cursor). Parameters: TOOLS.md.

Use as a dependency

npm install @shubhamcodess/tiny-browse

import { webSearch, webFetch } from "@shubhamcodess/tiny-browse";

const { results, pages, markdown } = await webSearch({ query: "node.js stream backpressure", read: 2 });
for (const r of results) console.log(r.rank, r.title, r.url);

const docs = await webFetch({ urls: ["https://go.dev/doc/effective_go"], focus: "goroutines", max_tokens: 2000 });
console.log(docs.pages[0].markdown);

Each call starts what it needs and shuts it down before returning, so short-lived CLIs exit straight away. Use createClient() to reuse one Chrome across calls, mode: "http" to skip Chrome entirely, and TinyBrowseError.code to handle failures. Full API: LIBRARY.md.

How it works

1. Search. Google's results page is loaded in your Chrome and read three independent ways (layout, accessibility tree, selectors), merged by vote, so a page redesign rarely breaks it. If Google is blocked or over its rate budget, DuckDuckGo and Bing are tried.

2. Read. Each page is fetched over plain HTTP and rendered with scripts off, which is fast; pages that need JavaScript get a full Chrome load. The faster way is remembered per site.

3. Extract. The main content is found by how the page is laid out, not by class names. Menus, ads, cookie banners and hidden text are dropped.

4. Fit. The page is split into sections by heading and packed into max_tokens, keeping the sections most relevant to the query.

Limitations

• Search engines may block automated browsers, depending on your network. Running tiny-browse login once usually helps with Google. Reading pages is not affected.

• Sites that refuse automated access with a bot check are reported as blocked, never bypassed.

• No PDF support yet.

• Tested on macOS and Linux. Windows is untested.

Troubleshooting

Documentation

License

MIT © Shubham Prakash