browser39

browser39

A headless browser for AI agents that fetches modern web pages, runs JavaScript, manages sessions, and returns token-efficient Markdown.

• Handles modern sites. Executes JavaScript, fills forms, queries the DOM, persists cookies and sessions across runs.
• LLM-usable output. Compact Markdown with content preselection, so the agent reads the section it needs, not the whole page.
• Local-only. No data sent to third-party services.
• Single binary. No Chrome, no Puppeteer. ~52MB. macOS, Linux, Windows.

Comparison

	browser39	Playwright / Puppeteer	Raw HTTP (requests, ureq)
External browser	None (single binary)	Requires Chrome/Chromium	None
Binary size	~52MB	~280MB with browser	N/A (library)
Platforms	macOS, Linux, Windows	macOS, Linux, Windows	Any
JavaScript	Yes (V8 via deno_core)	Yes (full V8)	No
HTML to Markdown	Built-in, token-optimized	No (raw HTML or screenshots)	DIY
Token preselection	Content sections, agent picks what to read	No	No
Cookies & sessions	Automatic, persisted, encrypted	Manual	Manual
DOM queries	CSS selectors + full JS DOM API	Full DOM API	No
Forms	fill + submit	Full interaction	Manual POST
Auth & secrets	Profiles, redaction, opaque handles	Manual	Manual
Transports	MCP (stdio + HTTP), JSONL, CLI	Library API	Library API

Token savings in practice

Real test: extracting the "Optical communications" section from Artemis II on Wikipedia (full page: ~14,600 tokens).

	Raw HTTP	WebFetch (Claude Code built-in)	Mistral Web Search	browser39
How it works	Fetch full page, truncate to ~1,000 tokens	Send full page (~14,600 tokens) to intermediate model with extraction prompt	Cloud API: search + page processing by Mistral model	Fetch → content selectors list → targeted section fetch
Tokens consumed	~1,000 (truncated)	~14,600 (processed by intermediate model)	Cloud processed, not disclosed	196
Found the section?	No. Section is at token ~6,320, truncated away	Yes, but returns a lossy summary	Depends on search ranking	Yes. Exact original content
Content quality	Nav menus, infobox, article intro	Paraphrased, no links, no references	Summary with citations	Lossless markdown with links and citations
Session state	None	None	None	Cookies, history, follow-up queries free
Data processing	Local	Processed remotely	Processed remotely	Local
Cost per call	Free	Bundled	$30 / 1,000 calls	Free
Retries needed	Pagination to find it	None, but no control over output	May not find specific section	None. Agent sees structure first

browser39 returns the exact section in 196 tokens at zero cost. The raw approach misses it entirely, WebFetch burns 75x more tokens through an intermediate model, and cloud tools like Mistral's charge $0.03 per call.

Install

npm install @aquintanar/browser39

Or via Cargo:

cargo install browser39

Install for any AI CLI / IDE

Installs the binary and auto-configures it for every MCP client detected: Claude Code, Claude Desktop, Codex, OpenCode, OpenClaw.

curl -fsSL https://raw.githubusercontent.com/alejandroqh/marketplace/main/h39.sh | bash

Pre-built binaries available on the releases page.

Quick Start

MCP config

Add to your MCP client config:

{
  "mcpServers": {
    "browser39": {
      "command": "browser39",
      "args": ["mcp"]
    }
  }
}

29 tools available instantly: browser39_fetch, browser39_click, browser39_links, browser39_dom_query, browser39_fill, browser39_submit, browser39_search, cookies, storage, history, config management, and more.

CLI: one-shot fetch

browser39 fetch https://example.com

# Example Domain
This domain is for use in documentation examples without needing permission.

[Learn more](https://iana.org/domains/example)

CLI: agent integration (watch mode)

Long-running subprocess that any language can talk to via JSONL files:

touch commands.jsonl
browser39 watch commands.jsonl --output results.jsonl

# From your agent (Python, Node, Rust, shell, anything):
echo '{"id":"1","action":"fetch","v":1,"seq":1,"url":"https://example.com"}' >> commands.jsonl