hydro13/tandem-browser

What it solves

Tandem Browser addresses the limitation of AI "sidebars" by integrating the AI agent directly into the browser runtime. This allows agents to operate within the user's real, authenticated sessions (cookies, logins) and interact with the web as it exists today, enabling the automation of SaaS platforms that lack APIs and the ability to modify live website UIs in real-time.

How it works

It is a local-first Electron browser that provides a first-class agent layer. Agents connect to the browser via the Model Context Protocol (MCP) or a comprehensive HTTP API (300+ endpoints). The agent interacts with the page using the accessibility tree, DOM, and live network logs rather than fragile pixel-based automation. To ensure safety, it employs an 8-layer security perimeter to prevent prompt injections and protect the user's machine from malicious web content.

Who it’s for

• AI Developers: Those building agents that need to perform complex web tasks without writing site-specific recipes.
• Power Users: People who want to automate their daily SaaS workflows (e.g., Gmail, banking portals) using their own logged-in sessions.
• Security-Conscious Users: Users who want a local-first browser where data and sessions remain on their own machine.

Highlights

• Model Agnostic: Compatible with any agent that speaks MCP or HTTP, including Claude, GPT, Gemini, and local models via Ollama or LM Studio.
• Real Session Access: Operates on the user's actual tabs and cookies, avoiding the need for headless browser setups or manual login scripts.
• Semantic Interaction: Uses accessibility snapshots and the DOM to "see" the page structure, making it more robust than screenshot-driven automation.
• Human-in-the-Loop: Supports explicit handoffs, allowing humans to handle CAPTCHAs or risky actions before the agent resumes.
• Remote Connectivity: Supports connecting agents remotely over a private Tailscale network.