Bindu - A2A Protocol Compliant AI Agent Framework

Real-browser automation agent powered by Notte — navigates JS-rendered pages, fills forms, solves captchas, and uses proxies.

Code

Create notte-browser-agent.py with the code below, or save it directly from your editor.

"""Notte Browser Agent (Bindu example).

Wraps the Notte SDK's built-in Agent + Session as a Bindu microservice.
Notte provides the browser runtime (cloud Chromium, stealth, captcha solving,
proxies, Vault-backed auth) and the agent loop that turns a natural-language
task into structured Pydantic output. `bindufy(...)` exposes that as a
Bindu-native agent with DID identity, A2A protocol, and skill routing.

Unlike most Bindu examples, there is no separate "LLM + tools" layer here —
Notte IS the agent. The handler simply forwards the user message to
`client.Agent(...).run(task=...)`.

Usage:
    python notte_browser_agent.py

Environment:
    Requires NOTTE_API_KEY in .env (https://console.notte.cc to obtain one).

Docs:
    - Notte SDK:            https://github.com/nottelabs/notte
    - Notte docs:           https://docs.notte.cc
    - Agent Skill (Claude): https://github.com/nottelabs/agent-skill-notte
"""

import os

from dotenv import load_dotenv

load_dotenv()

from bindu.penguin.bindufy import bindufy
from notte_sdk import NotteClient

# A single, long-lived Notte client. Session lifetime is scoped per request
# inside the handler so every A2A call gets an isolated browser context.
client = NotteClient()

# Sensible default model. Escalate to anthropic/claude-sonnet-4-5 or
# openai/gpt-4.1 for harder multi-step flows via NOTTE_REASONING_MODEL.
REASONING_MODEL = os.getenv("NOTTE_REASONING_MODEL", "gemini/gemini-2.5-flash")
MAX_STEPS = int(os.getenv("NOTTE_MAX_STEPS", "15"))
SOLVE_CAPTCHAS = os.getenv("NOTTE_SOLVE_CAPTCHAS", "false").lower() == "true"
USE_PROXIES = os.getenv("NOTTE_USE_PROXIES", "false").lower() == "true"

# Bindu agent configuration
config = {
    "author": "bindu.builder@getbindu.com",
    "name": "notte_browser_agent",
    "description": (
        "Real-browser web automation agent powered by Notte. Navigates JS-rendered "
        "pages, fills forms, handles auth, solves captchas, and returns Pydantic-"
        "validated structured output."
    ),
    "deployment": {
        "url": os.getenv("BINDU_DEPLOYMENT_URL", "http://localhost:3773"),
        "expose": True,
        "cors_origins": ["http://localhost:5173"],
    },
    "skills": ["skills/notte-browser-skill"],
}


def handler(messages: list[dict[str, str]]) -> str:
    """Run the latest user message as a Notte browser task.

    Args:
        messages: A2A message history; we use the latest user turn as the task.

    Returns:
        The agent's final answer as a string (JSON if the caller asked for
        structured output, free text otherwise).
    """
    if not messages:
        return (
            "Please provide a browser task. Example: 'Go to news.ycombinator.com "
            "and return the top 5 posts as JSON with title, url, and points.'"
        )

    latest_user = next(
        (
            m
            for m in reversed(messages)
            if isinstance(m, dict) and m.get("role") == "user"
        ),
        None,
    )
    if latest_user is None:
        return "No user message found — please send a task as a user-role message."
    task = latest_user.get("content", "")
    if not isinstance(task, str) or not task.strip():
        return "Empty task — please describe the browser action you want."

    with client.Session(
        solve_captchas=SOLVE_CAPTCHAS,
        proxies=USE_PROXIES,
    ) as session:
        agent = client.Agent(
            session=session,
            reasoning_model=REASONING_MODEL,
            max_steps=MAX_STEPS,
        )
        response = agent.run(task=task)

    answer = getattr(response, "answer", None)
    if answer is None:
        return "Notte agent returned no answer — try raising NOTTE_MAX_STEPS or a stronger NOTTE_REASONING_MODEL."
    return answer


if __name__ == "__main__":
    bindufy(config, handler)

Skill Configuration

Create skills/notte-browser-skill/skill.yaml:

id: notte-browser-skill
name: notte-browser-skill
version: 1.0.0
author: bindu.builder@getbindu.com
description: |
  Real-browser automation skill powered by Notte. Navigates JavaScript-rendered
  pages, fills forms, clicks through multi-step flows, solves captchas, and
  routes through proxies. Use whenever the task needs a real browser (not just
  HTTP) — dynamic pages, e-commerce flows, or extraction from pages that won't
  respond to `requests` / `curl`.

tags:
  - browser-automation
  - web-agent
  - web-scraping
  - form-filling
  - captcha
  - proxies

input_modes:
  - application/json

output_modes:
  - application/json
  - text/plain

examples:
  - "Go to news.ycombinator.com and return the top 5 posts as JSON with title, url, points, author"
  - "Search 'air force 1' on nike.com, solve any captcha, and return the current price and availability"
  - "Fill the contact form at https://example.com/contact with name=Jane, email=jane@ex.com, submit, and confirm the success state"

capabilities_detail:
  real_browser:
    supported: true
    description: "Runs a Chromium-based Notte cloud session — JavaScript renders, forms submit, cookies persist. Local or alternate-browser runtimes require passing cdp_url / extra params to client.Session()."
  multi_step_workflows:
    supported: true
    description: "Natural-language task routed through a Notte agent with configurable max_steps and reasoning_model"
  captcha_solving:
    supported: true
    description: "Set NOTTE_SOLVE_CAPTCHAS=true to enable automatic captcha resolution on the Session"
  proxies:
    supported: true
    description: "Set NOTTE_USE_PROXIES=true to route traffic through Notte's managed proxies"

assessment:
  keywords:
    - browser
    - navigate
    - click
    - fill
    - form
    - login
    - captcha
    - proxy
    - dynamic
    - javascript
    - spa
    - multi-step
    - checkout
    - booking

  specializations:
    - domain: dynamic_page_extraction
      confidence_boost: 0.3
    - domain: multi_step_web_workflows
      confidence_boost: 0.3
    - domain: e_commerce_checkout
      confidence_boost: 0.25

  anti_patterns:
    - "static html fetch"
    - "public json api"
    - "offline file processing"
    - "pdf text extraction without a browser"
    - "database query"

How It Works

Notte SDK configuration

NotteClient(): a single long-lived client read from NOTTE_API_KEY in the environment.
client.Session(solve_captchas=..., proxies=...): a context-managed browser session, scoped per request so every A2A call gets an isolated Chromium context with its own cookies and fingerprint.
client.Agent(session, reasoning_model, max_steps): the agent loop on top of the session.
- reasoning_model (NOTTE_REASONING_MODEL, default gemini/gemini-2.5-flash) — LiteLLM model id; escalate to anthropic/claude-sonnet-4-5 or openai/gpt-4.1 for harder multi-step flows.
- max_steps (NOTTE_MAX_STEPS, default 15) — per-task step budget.
NOTTE_SOLVE_CAPTCHAS=true flips on Notte’s built-in captcha solver for the session.
NOTTE_USE_PROXIES=true routes traffic through Notte’s managed proxy pool, which helps with rate-limited or geo-blocked targets.

Bridging the bindu message → Notte task

The handler walks messages in reverse to find the latest role == "user" entry and pulls its content string. That string becomes the task argument to agent.run(...) — there’s no prompt construction, no tool catalog wiring, no system prompt to author. Notte is the agent.
Empty histories, missing user turns, and empty content all return helpful error strings rather than raising.

What agent.run() returns

A response object whose .answer attribute is a single string. If the task asks for structured output (“…as JSON with keys X, Y”) the string is JSON; otherwise it’s free text.
The handler returns response.answer as-is — no schema enforcement on the Bindu side. If .answer is None, the handler returns a hint to raise NOTTE_MAX_STEPS or use a stronger reasoning model.

Why use it

The headless browser runs in Notte’s cloud, not on your machine. JavaScript renders, forms submit, captchas can be solved, and proxies are available — all without you wiring Playwright, Selenium, or a captcha provider.

Dependencies

uv init
uv add bindu python-dotenv notte-sdk

notte-sdk isn’t bundled with Bindu’s agents extra — install it explicitly or boot fails on AuthenticationError.

Environment Setup

Create .env file:

# Notte API key — get one at https://console.notte.cc
NOTTE_API_KEY=your_notte_api_key

# Optional tuning
NOTTE_REASONING_MODEL=gemini/gemini-2.5-flash
NOTTE_MAX_STEPS=15
NOTTE_SOLVE_CAPTCHAS=false
NOTTE_USE_PROXIES=false

# Optional Bindu deployment override
BINDU_DEPLOYMENT_URL=http://localhost:3773

Run

uv run notte-browser-agent.py

Examples:

“Find the cheapest flight from SF to LA on Google Flights for next Monday”
“Go to news.ycombinator.com and return the top 5 posts as JSON with title, url, and points”
“Search ‘air force 1’ on nike.com, solve any captcha, and return the current price and availability”
“Fill the contact form at https://example.com/contact with name=Jane, email=jane@ex.com, submit, and confirm the success state”

Expect 10–30s per task for multi-step flows. Each request gets its own browser context.

Example API Calls

Message Send Request

{
  "jsonrpc": "2.0",
  "method": "message/send",
  "params": {
    "message": {
      "role": "user",
      "kind": "message",
      "messageId": "9f11c870-5616-49ad-b187-d93cbb100001",
      "contextId": "9f11c870-5616-49ad-b187-d93cbb100002",
      "taskId": "9f11c870-5616-49ad-b187-d93cbb100003",
      "parts": [
        {
          "kind": "text",
          "text": "Find the cheapest flight from SF to LA on Google Flights for next Monday and return the airline, departure time, and price as JSON."
        }
      ]
    },
    "skillId": "notte-browser-skill",
    "configuration": {
      "acceptedOutputModes": ["application/json"]
    }
  },
  "id": "9f11c870-5616-49ad-b187-d93cbb100003"
}

Task get Request

{
  "jsonrpc": "2.0",
  "method": "tasks/get",
  "params": {
    "taskId": "9f11c870-5616-49ad-b187-d93cbb100003"
  },
  "id": "9f11c870-5616-49ad-b187-d93cbb100004"
}

Frontend Setup

# Clone the Bindu repository
git clone https://github.com/GetBindu/Bindu

# Navigate to frontend directory
cd frontend

# Install dependencies
npm install

# Start frontend development server
npm run dev

Open http://localhost:5173 and try to chat with the Notte browser agent.

Beginner

Specialized

Advanced

2.12 Notte Browser Agent

Code

Skill Configuration

How It Works

Dependencies

Environment Setup

Run

Example API Calls

Frontend Setup

​Code

​Skill Configuration

​How It Works

​Dependencies

​Environment Setup

​Run

​Example API Calls

​Frontend Setup

Code

Skill Configuration

How It Works

Dependencies

Environment Setup

Run

Example API Calls

Frontend Setup