Show HN: I made web agents reliable with smaller LLMs via natural language

2 days ago 5

Notte provides the full-stack web AI agents framework that allows you to develop, deploy, and scale your own agents, all with a single API. Transform the internet into an agent-friendly environment and websites into structured, navigable maps described in natural language.

Browser Sessions → on-demand headless browser instances, built in & custom proxy config, CDP, cookie integration, session replay
Run automated LLM-powered agents → solve complex tasks on the web
Observe, Step, Scrape → observe website states and execute actions using intuitive natural language commands — granular control while maintaining the simplicity of natural language interaction
Secrets Vault → enterprise-grade credential management for your Sessions & Agents

🚀 Zero Friction Agentic Browsing

notte.agents.run(task="go to twitter and post: new era this is @nottecore taking over my acc")

Fast, reliable, agentic browsing (speed x2):

secure password vault
bypass bot detection

Rank Provider Agent Self-Report LLM Evaluation Time per Task Task Reliability

🏆	Notte	86.2%	79.0%	47s	96.6%
2️⃣	Browser-Use	77.3%	60.2%	113s	83.3%
3️⃣	Convergence	38.4%	31.4%	83s	50%

Read the full story here: https://github.com/nottelabs/open-operator-evals

To run the above you'll need a notte API key from our console platform 🔑

uv venv --python 3.11 uv pip install notte-sdk

...and spin up your crazy cool and dead simple agent on our remote machine:

import os from notte_sdk import NotteClient notte = NotteClient(api_key=os.getenv("NOTTE_API_KEY")) agi = notte.Agent(headless=False,reasoning_model="gemini/gemini-2.0-flash", max_steps=5) agi.run(task="doom scroll cat memes on google images")

This is by far the closest attempt to AGI we've ever witnessed ;)

Our main tech highlight is that we introduce a perception layer that turns the internet into an agent-friendly environment, by turning websites into structured maps described in natural language, ready to be digested by an LLM with less effort.

from notte_sdk import NotteClient notte = NotteClient() with notte.Session(headless=False) as page: obs = page.observe(url="https://www.google.com/travel/flights") print(obs.space.markdown)

# Flight Search * I1: Enters departure location (departureLocation: str = "San Francisco") * I3: Selects departure date (departureDate: date) * I6: Selects trip type (tripType: str = "round-trip", allowed=["round-trip", "one-way", "multi-city"]) * B3: Search flights options with current filters # Website Navigation * B5: Opens Google apps menu * L28: Navigates to Google homepage # User Preferences * B26: Open menu to change language settings ...

The above gives you the gist of how we push to better parse webpages and reduce the cognitive load of LLM reasoners. The aim is to enable you to build and deploy more accurate web browsing agents, while downgrading to smaller models, which in turn increase inference speed and reduce production costs.

☄️ Speed contest vs. Browser-Use

The perception layer enables smaller models (e.g. the llama suite) to be connected for the agent's reasoning, because all the DOM noise is abstracted and the LLM can focus on a set of actions described in plain language. This allows the agent to be served on ultra-high inference such as Cerebras without losing precision:

from notte_sdk import NotteClient notte = NotteClient() agent = notte.Agent(headless=False,reasoning_model="gemini/gemini-2.0-flash", max_steps=5) agent.run(task="search cheapest flight from paris to nyc on gflight")

left:browser-use, right:notte-agent (cerebras)

🌐 The Full Stack Framework

Notte's full stack agentic internet framework combines core browser infrastructure (sessions, live replay, cdp) with intelligent browsing agents, bridged and enhanced with our perception layer. Our entire codebase is made to be highly customizable, ready to integrate other devtools from the ecosystem and packaged to be push to prod. We also provide web scripting capabilities and sota scraping endpoints out of the box, because why not.

service agent.run() agent.cloud() page.scrape() page.step() page.perceive()

browser-use	🌕	🌕
stagehand			🌕	🌕
notte	🌕	🌕	🌕	🌕	🌕

PS: The title of services are figurative eg. agent.cloud() refers to hosting an agent in cloud for you.

Unstable and upcoming features

⏭️ We have either already partially shipped or are working on the following features: captcha resolution, residential proxies, web security, vpn-style browsing, authentication and payments with secure safe, improved speed and memory, human-in-the-loop integration, channeled notifications, and cookies management.

Scraping endpoint:

/v1/scrape - Scrape data from a URL

Session management:

/v1/sessions/start - Create a new browser session
/v1/sessions/{session_id}/stop - Close a session
/v1/sessions/{session_id}/replay - Get replay information (video in .webp format)
/v1/sessions/{session_id}/debug - Get debugging information from a session (i.e live CDP url / viewer url)
/v1/sessions - List active sessions

Browser & Page interactions:

/v1/sessions/{session_id}/page/scrape - Extract structured data from current page
/v1/sessions/{session_id}/page/observe - Get action space (perception) from current page
/v1/sessions/{session_id}/page/step - Perform action on current page with text command

Agent launchpad:

/v1/agent/start - Start a new agent task ()
/v1/agent/{agent_id} - Get agent task status
/v1/agent/{agent_id}/stop - Stop running agent
/v1/agents/ - List running agent tasks

Read more on our documentation website. You can cURL all of them 🥰

You use our [console]((https://console.notte.cc) to manage your agents and sessions, create API keys and monitor your usage.

Most of our open-source features are also available on the cloud-hosted version with a large free-tier!

Example for webpage scraping & structured schema llm extraction

from notte_sdk import NotteClient notte = NotteClient() data = notte.scrape(url="https://pump.fun", instructions="get top 5 latest trendy coins on pf, return ticker, name, mcap")

How to run notte locally ?

You will need to provide your own LLM provider API key, and install the dependencies:

uv venv --python 3.11 uv pip install notte uv run patchright install --with-deps chromium export GEMINI_API_KEY="your-api-key"

...and spin up your crazy cool and dead simple agent on your local machine:

import notte agi = notte.Agent(headless=False,reasoning_model="gemini/gemini-2.0-flash", max_steps=5) agi.run(task="doom scroll cat memes on google images")

Setup your local working environment;

uv sync --all-extras --dev uv run patchright install --with-deps chromium uv run pre-commit install

Find an issue, fork, open a PR, and merge :)

This project is licensed under the Server Side Public License v1. See the LICENSE file for details.

If you use notte in your research or project, please cite:

@software{notte2025, author = {Pinto, Andrea and Giordano, Lucas and {nottelabs-team}}, title = {Notte: Software suite for internet-native agentic systems}, url = {https://github.com/nottelabs/notte}, year = {2025}, publisher = {GitHub}, license = {SSPL-1.0} version = {1.4.4}, }

Built with luv from Earth 🌏

Read Entire Article