bladebro

Give your AI agent a browser. Few tools. Full control. Real stealth. Zero runtime deps.

Re-render-immune refs · batch actions · auto-extract · self-improving · 6-layer stealth

One MCP server · one persistent page model · zero Node.js · one binary · Linux · macOS · Windows

npm install -g bladebro && bladebro mcp

Install · Stealth Bench V1 · The 5 tools · Architecture · Re-render immunity · Self-improvement · Stealth · Comparison · Gotchas · Limits

Bladebro is an agentic browser driver built from the agent's perspective. Instead of 20+ tools that each do one thing, Bladebro gives you 5 tools that together provide full control. It drives stock Chromium over CDP, holds a persistent Live Page Model across tool calls, and returns diff-first results: the agent sees what changed, not the whole world, every single time.

🏆 Stealth Bench V1: 85% (68 / 80)

Bladebro was run against browser-use's Stealth Bench V1: 80 real, anti-bot-protected production sites across 11 vendors. It walked through 68 of 80 (85%), above every provider on browser-use's published leaderboard (browser-use-cloud 81%).

• Out of the box: one IP, no proxy, no captcha solver, no fingerprint rotation, no cloud. Stock Chromium with Bladebro's stealth layer on by default.
• Honest run: real navigation of the real 80 sites; an opencode agent (open model) acted as the browsing agent because we had no paid access to browser-use's cloud agent, and the deviations are disclosed in full. The benchmark isolates browser stealth from agent ability, and the defensive layer being measured is Bladebro's own.
• Per-vendor: perfect on Cloudflare (22/22) and reCaptcha (6/6); cleared the two sites browser-use-cloud's own results did not (Shape, Temu). Single named gap: Akamai (3/6).

Full report and methodology → · All 80 per-site results (CSV)

Demo

Built in Rust. One static binary. No runtime. No Node.js. No Playwright shim. Just the browser engine and your agent. Native on Linux, macOS, and Windows.

Speaks MCP 2024-11-05 through 2026-07-28: legacy initialize handshake and the new stateless per-request negotiation with server/discover, dual-dialect. Works with every MCP client, old and new.

What makes it different

🚀 Install

npm (recommended)

npm install -g bladebro
bladebro mcp

That's it. No Rust, no compilation, no dependencies. The npm package ships a prebuilt binary for your platform:

npm resolves the correct binary automatically via os/cpu fields. Users only download the binary for their platform. Zero postinstall scripts, zero warnings.

Pi coding agent

pi install npm:bladebro

That's it. The extension spawns the binary as a stdio MCP subprocess, discovers tools via tools/list, and registers them natively with pi via pi.registerTool(). The agent gets 5 first-class tools (browser.act, browser.see, browser.state, browser.run, browser.vision) — no adapter, no config files, no proxy tool.

Tool definitions come from the binary at startup, so they auto-adapt to any tool def changes with zero extension maintenance. Auto-updates via pi update --extensions.

From source

Prerequisites: Chromium or Google Chrome (auto-detected), Rust 1.86+. Linux: Xvfb for headless servers (macOS/Windows run headful natively).

git clone https://github.com/dondai44423/bladebro.git
cd bladebro
cargo build --release
./target/release/bladebro mcp

🔌 Two ways to use Bladebro

Bladebro gives you the same 5 tools, same stealth, same page model through two interfaces. Pick one or use both.

Option 1: MCP Server

Just tell your agent to add Bladebro to its MCP config. For Claude Desktop, Cursor, and most MCP clients, add this to your config file:

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

Using pi? One command, no config:

pi install npm:bladebro

Option 2: CLI

The CLI has the exact same power as the MCP server. Same handlers, same stealth, same page model. Any feature update auto-propagates to both surfaces automatically.

How AI agents discover the CLI: bladebro help --json returns the same structured tool definitions as MCP tools/list, plus a CLI command mapping. An agent calls it once to learn the full interface, then uses --json on every command for structured output. No guessing, no parsing help text.

# Agent discovery: same schemas as MCP tools/list
bladebro help --json | jq '.tools[].name'
# ["act", "see", "state", "run", "vision"]

Daemon mode (persistent Chrome, zero startup delay after first launch):

# Start the daemon (Chrome stays alive across commands)
bladebro daemon

# All commands now connect to the daemon instead of launching new Chrome
bladebro nav https://news.ycombinator.com
bladebro see content
bladebro act click e5
bladebro see model --json | jq .text
bladebro state cookies
bladebro vision --marks
bladebro stop

One-shot mode (no daemon, launches Chrome per command):

bladebro see content https://example.com --no-daemon
bladebro nav https://news.ycombinator.com --no-daemon

All 5 tools work from the CLI:

# Navigate
bladebro nav https://example.com

# Read the page (6 modes: model, content, outline, extract, links, forms)
bladebro see model                   # interactive elements with refs
bladebro see content                # clean markdown
bladebro see outline                 # heading hierarchy
bladebro see extract auto            # auto-detect structured data

# Interact (20+ actions: click, type, fill, scroll, press, hover, ...)
bladebro act click e5                # click element e5
bladebro act type e12 "hello world"  # type text
bladebro act scroll 0 500            # scroll down
bladebro act press Enter             # press