peerd is a novel AI agent harness designed to run entirely within the user's browser as a Chrome or Firefox extension. It enables users to execute a full agent loop, interact with browser tabs, and leverage sandboxed compute environments like JavaScript notebooks and WebAssembly-based Linux VMs. The primary benefit is a backend-free, telemetry-free architecture that respects user privacy and allows for Bring Your Own Key (BYOK) model integration, all while utilizing the browser's inherent security model.

How It Works

peerd operates by running a complete agent loop directly within the browser, building upon decades of hardened browser platform security. It employs a unique architecture where a main orchestrator agent delegates tasks to specialized "actor" sub-agents. Each actor is confined to a specific environment (e.g., a browser tab, a sandboxed VM, or a notebook) and exclusively holds that environment's tools, operating without direct access to the orchestrator's keys. This isolation ensures that the agent holding sensitive keys never directly interacts with raw page content or executes untrusted code. Furthermore, every action taken by an actor is verified against the live DOM before being considered complete, creating a robust security boundary.

Quick Start & Requirements

• Primary install / run command: Clone the repository (git clone https://github.com/NotASithLord/peerd.git), cd peerd, and load the extension/ directory unpacked in a Chromium-based browser (Chrome, Edge, Brave, Arc) via chrome://extensions with Developer mode enabled. For Firefox, install the peerd-preview-firefox.xpi from GitHub Releases.
• Non-default prerequisites and dependencies: A Chromium-based browser or Firefox. API keys for model providers like Anthropic, OpenRouter, or a local Ollama instance are required.
• Estimated setup time or resource footprint: Initial setup involves cloning the repo and loading the extension, followed by configuring the local vault and API keys. No complex build steps are required for basic operation.
• Links: Developer preview installation instructions are detailed in the README. GitHub Releases for preview packages.

Highlighted Details

• Browser-Native Agent: Runs a full AI agent loop directly within Chrome/Firefox, leveraging existing tabs and sessions.
• Sandboxed Compute: Supports JavaScript Notebooks, WebAssembly Linux VMs (CheerpX-emulated Debian), and client-side applications within secure, isolated browser environments.
• Peer-to-Peer Sharing: (Preview channel) Enables agent-to-agent communication over a WebRTC network.
• BYOK & Local Vault: Users provide their own model API keys, which are encrypted and stored securely in a local vault unlocked via passkey or passphrase.
• Enhanced Security: Employs browser security primitives (V8 isolates, WebCrypto, WebAuthn) and an actor-based delegation model to prevent the main agent from accessing raw data or executing arbitrary code.

Maintenance & Community

The project is explicitly marked as "0.x, experimental beta," with a warning that breaking changes are likely and the surface area is still evolving. There are no explicit mentions of a dedicated community forum (like Discord/Slack), roadmap, or notable contributors within the provided README. The code itself is presented as the specification.

Licensing & Compatibility

• License type: The core project is licensed under Apache 2.0.
• Compatibility notes: Designed for Chromium-based browsers and Firefox. However, a critical caveat exists regarding the CheerpX component, which is proprietary software. While the free Community tier may cover individual use, organizations with more than one person or those wishing to self-host the runtime require a paid Commercial License from Leaning Technologies. Google's Gemma model weights also have custom terms of use with restrictions.

Limitations & Caveats

peerd is in an experimental beta stage (0.x), meaning breaking changes are expected, and the API surface is unstable. The most significant adoption blocker for commercial or organizational use is the proprietary license for CheerpX, which is essential for the WebAssembly Linux VM functionality. Additionally, the Gemma model weights used for on-device inference are subject to Google's custom terms of use, which are not a standard OSI-approved license.