← Back to Projects

Agentic Browser

A model-agnostic agentic browser powered by a Python MCP server and WebExtensions API — enabling AI-driven reasoning, form filling, RAG workflows, and transparent, human-supervised web automation.

Python
View Source

Agentic Browser — Adaptive, Model-Agnostic Web Automation

The Open Agent Browser Extension Powered by Python & MCP

Mission: Build an intelligent browser agent that doesn’t just understand the web — it acts on it. Fully model-agnostic, privacy-respecting, and BYOKeys-ready.

Overview

Agentic Browser is a next-generation browser extension powered by a Python MCP (Model Context Protocol) server that bridges modern LLM reasoning with real browser interactivity.

Unlike typical AI assistants, this agent:

  • Understands complex web content,
  • Takes actions (like filling forms, navigating, comparing data),
  • And adapts to any preferred model backend — OpenAI, Anthropic, Ollama, local LLaMA, Mistral, etc.

It’s your agent, your browser, your keys.

Architecture

image

Key Architectural Principles

  • Model-Agnostic: Works with any LLM backend that supports API-style calls (OpenAI-compatible, Anthropic, Ollama, LM Studio).
  • BYOKeys: No vendor lock-in. Users supply their own API keys via .env or runtime UI input.
  • MCP-Compliant: Uses the Model Context Protocol for secure and structured interaction.
  • Declarative Action System: The model declares browser actions (e.g. click, fill_form, extract), and the extension executes them safely.

Core Objectives

1. Model-Agnostic Agent Backend

Create a flexible, LLM-agnostic backend using Python, LangChain, and the Model Context Protocol (MCP) framework.
Allows seamless switching across models (OpenRouter, Ollama, Anthropic, OpenAI, or local inference models).

2. Secure Browser Extension

Design a robust and secure browser extension using the WebExtensions API, ensuring compatibility across Chrome, Firefox, and other Chromium-based browsers.

3. Advanced Agent Workflows

Support sophisticated agentic workflows through Retrieval Augmented Generation (RAG), persistent memory, and automated multi-step browsing tasks like form filling, search synthesis, and citation retrieval.

4. Guardrails & Transparency

Implement strong security and transparency layers:

  • User approval before every actionable operation
  • Comprehensive activity logs
  • Intelligent content filtering
  • Safe domain allowlisting and IPI protection

5. Open-Source Extensibility

Adopt a modular, community-driven architecture encouraging open innovation and integration of new capabilities, workflows, and extensions over time.

Technical Stack: Components & Technologies

Component Functionality Technologies / Frameworks
Agent Orchestration Task planning, retrieval-augmented reasoning, complex multi-step workflows LangChain, LangGraph
Browser Control DOM inspection, navigation, form filling, input injection, and content extraction WebExtensions API (Chrome / Firefox)
LLM Adapters Model-agnostic routing, adapter layer for multi-provider compatibility OpenRouter, Ollama, Anthropic, OpenAI, Hugging Face APIs
Backend Agent Core logic execution, action orchestration, safety and state management Python MCP Server
Retrieval & Citation Web data extraction, embedding-based retrieval, factual grounding Vector Databases (FAISS / Pinecone)
Safety & Guardrails Logging, data protection, domain-level security enforcement Secure Audit System, Activity Logger

Core Features

Model-Agnostic Intelligence

Works with any LLM provider — OpenAI, Anthropic, Mistral, Ollama, LM Studio, or custom deployments — using a unified adapter layer.

Bring Your Own Keys (BYOKeys)

No vendor dependency. Users supply their own API keys securely via local .env or UI input; keys never leave local context.

Web Interaction Engine

Real-time DOM inspection and manipulation for safe, human-approved automation — including form filling, data extraction, and structured web actions.

Retrieval & Grounded Reasoning

Leverages RAG pipelines to incorporate external data and enhance factual grounding, improving contextual accuracy in responses.

Secure Architecture

Every action is validated, logged, and requires explicit permission, ensuring responsible automation and explainability.

Extensible Agent Tools

Developers can easily extend agent capabilities by adding Python tools, context managers, or new browser-side actions.

Roadmap

  • Add visual DOM debugger panel
  • Multi-model round-robin support (reasoning blending)
  • Offline LLM embedding-based retrieval
  • GUI for managing keys/providers
  • Fine-grained content permissions

Contributing

Contributions are very welcome!
If you’re into LLM orchestration, WebExtension APIs, or intelligent web automation, this project is an open canvas.

Please:

  1. Fork the repo
  2. Create a feature branch
  3. Submit a well-documented PR

License

Released under the GPL 3 License — free to modify, distribute, and extend with attribution.