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: one-shot search

Search the web using the configured search engine:

browser39 search "rust async traits"

Add --output json for structured results suitable for piping into other tools.

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

Drop-in web_search and visit_website tool examples: Python | TypeScript | Rust

See docs/install-cli.md for the full integration guide.

Rust library

Embed browser39 directly in a Rust application — no subprocess, no IPC.

[dependencies]
browser39 = "1.8"
tokio = { version = "1", features = ["full"] }
anyhow = "1"

use std::collections::HashMap;

use browser39::{BrowserService, Config, HttpMethod, InMemoryStore, SessionStore};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let store: Box<dyn SessionStore> = Box::new(InMemoryStore);
    let mut service = BrowserService::new(Config::default(), store).await?;

    let options = service.default_fetch_options();
    let page = service
        .fetch(
            "https://example.com",
            &HttpMethod::Get,
            &HashMap::new(),
            None,
            None,
            &options,
        )
        .await?;

    println!("{}", page.markdown);
    Ok(())
}

The crate root re-exports the common types: BrowserService, Config, PersistenceMode, FetchOptions, HttpMethod, SessionStore, InMemoryStore, create_session_store. Lower-level modules (browser39::core, browser39::cli, browser39::mcp) are also public if you need direct access to the HTML→Markdown engine, JSONL runner, or MCP server.

Features

Token optimization

browser39 minimizes token usage when feeding web content to LLMs:

• Content preselection: on first fetch, returns available content sections with token estimates instead of dumping the full page. The agent picks the relevant section and re-fetches with a targeted selector.
• Heading auto-expand: selector: "#Astronauts" returns the full section until the next same-level heading, not just the heading text.
• HTML to Markdown: strips scripts, styles, and non-content elements.
• Compact link references (JSON mode): [text][N] instead of inline URLs, with full URLs in the links array.
• Same-origin URL shortening: links on the same domain show path-only.
• Link deduplication: same-URL links (image + headline cards) emitted once.

JavaScript execution

V8 (via deno_core) runs JavaScript against a full DOM environment:

• Traversal: parentElement, children, firstChild, lastChild, nextSibling, previousSibling, closest(), matches(), contains()
• Lookup: getElementById, getElementsByClassName, getElementsByTagName, getElementsByName
• Mutation: createElement, createTextNode, appendChild, removeChild, insertBefore, setAttribute, removeAttribute, textContent/innerHTML setters
• Events: addEventListener, removeEventListener, dispatchEvent, new Event/CustomEvent/MouseEvent/KeyboardEvent/InputEvent
• Web APIs: localStorage, document.cookie, console.log (captured), setTimeout, atob/btoa, getComputedStyle, MutationObserver
• Forms: element.value get/set, element.click(), form.submit()

{"action": "dom_query", "script": "document.querySelectorAll('a').length"}
{"action": "dom_query", "script": "document.getElementById('content').closest('section').textContent"}
{"action": "dom_query", "script": "document.querySelector('h1').setAttribute('class', 'modified')"}

Session persistence

Cookies, localStorage, and browsing history are persisted to disk by default (~/.local/share/browser39/session.enc, AES-256-GCM encrypted). An agent can log in once and stay authenticated across restarts.

Disable with --no-persist or config:

[session]
persistence = "memory"

Forms

Fill fields by CSS selector and submit. browser39 handles enctype, builds the HTTP request, and returns the response page:

{"action": "fill", "fields": [{"selector": "#user", "value": "agent"}, {"selector": "#pass", "value": "secret", "sensitive": true}]}
{"action": "submit", "selector": "form#login"}

Security

Auth profiles keep credentials out of the LLM conversation. The agent references a profile name and never sees the token:

[auth.github]
header = "Authorization"
value_env = "GITHUB_TOKEN"
value_prefix = "Bearer "
domains = ["api.github.com"]

{"action": "fetch", "url": "https://api.github.com/repos", "auth_profile": "github"}

Config management via MCP

Agents can manage browser39's configuration directly through MCP tools: change the search engine, store credentials, manage auth profiles, cookies, storage, and headers. Sensitive values are stored securely on disk but never returned via MCP; config_show masks them with ••••••.

> browser39_config_set key="search.engine" value="https://www.google.com/search?q={}"
Set search.engine = https://www.google.com/search?q={}

> browser39_config_auth_set name="github" header="Authorization" value="Bearer ghp_..." domains=["api.github.com"]
Auth profile 'github' saved

> browser39_config_show section="auth"
{"auth": {"github": {"header": "Authorization", "value": "••••••", ...}}}

10 config tools: config_show, config_set, config_auth_set/delete, config_cookie_set/delete, config_storage_set/delete, config_header_set/delete.

All transports

Transport	Command	Use case
MCP (stdio)	browser39 mcp	Local MCP clients
MCP (HTTP)	browser39 mcp --transport sse --port 8039	Remote agents, cloud deployments
JSONL watch	browser39 watch commands.jsonl	Any language, long-running agent IPC
JSONL batch	browser39 batch commands.jsonl	One-shot scripted operations
CLI fetch	browser39 fetch <url>	Quick page retrieval, shell scripts
CLI search	browser39 search <query>	Quick web search from shell scripts
Rust library	cargo add browser39	Embed in Rust apps, no subprocess

Configuration

browser39 --config path/to/config.toml fetch https://example.com

Precedence: --config flag > BROWSER39_CONFIG env > ~/.config/browser39/config.toml

See docs/config.md for the full reference.

Documentation

Doc	Description
install-cli.md	CLI integration guide with Rust, Python, TypeScript examples
jsonl-protocol.md	Full JSONL protocol specification
config.md	Configuration reference

Development

cargo build              # Build
cargo run                # Run
cargo test               # Run all tests
cargo clippy             # Lint
cargo fmt                # Format

Contributors

@nathan-widjaja

@janpauldahlke

License

Apache-2.0