peerd

peerd runs the agent loop directly inside the browser as a Manifest V3 extension, rather than requiring a separate backend server to orchestrate the agent. It can drive your open tabs, spin up sandboxed compute environments — JS notebooks and even WebAssembly-based Linux VMs — and share whatever it builds peer-to-peer with other users via WebRTC, without routing anything through a central server.

The project is explicitly Bring Your Own Key (BYOK): no backend and no telemetry means your model API key and agent activity never pass through the project’s own infrastructure, since there is none. It’s built with vanilla JavaScript and no build step, which the project treats as a deliberate design constraint rather than a limitation.

Apache-2.0 licensed and explicitly labeled 0.x experimental, peerd targets both Chrome and Firefox via the Manifest V3 extension standard, with a security policy and CI-gated releases in place despite its early-stage status.

What You Get

• An AI agent loop running natively inside a Manifest V3 browser extension, with no separate backend server
• Sandboxed compute environments spun up on demand — JS notebooks and WebAssembly-based Linux VMs
• Peer-to-peer sharing of agent-built artifacts via WebRTC, without routing through a central server
• BYOK model access with no telemetry, since there’s no backend to send data to

Common Use Cases

• Running an AI agent that drives browser tabs directly without a separate server-side orchestration layer
• Spinning up sandboxed compute (notebooks or WASM VMs) inside the browser for agent-driven experimentation
• Sharing agent-built artifacts peer-to-peer with other users without a central hosting service
• Running agent workflows with strict privacy requirements, since there’s no backend to route data through

Under The Hood

Architecture peerd’s defining architectural choice is running the entire agent harness inside the browser extension process itself — no external server coordinates the agent loop, and WebRTC handles peer-to-peer artifact sharing directly between browsers rather than through a relay server. Sandboxed compute (JS notebooks, WASM-based Linux VMs) runs within the browser’s own sandboxing model, letting the agent execute code without a separate compute backend.

Tech Stack Vanilla JavaScript with no build step, targeting Manifest V3 for both Chrome and Firefox extension compatibility, WebAssembly for in-browser Linux VM sandboxing, and WebRTC for peer-to-peer data sharing. The project maintains 100% ts-check type coverage despite being plain JavaScript, using TypeScript’s type-checking against JS via JSDoc-style annotations.

Code Quality CI-gated package-and-release workflows, a documented security policy, and enforced type-check coverage are unusually mature process signals for a project explicitly labeled “0.x experimental” — suggesting the author is treating engineering discipline as a priority even during early, rapid iteration.

What Makes It Unique Most AI agent tools depend on a backend service for orchestration, model routing, or artifact storage; peerd’s zero-backend, zero-telemetry, browser-native design with WebRTC-based P2P sharing is a structurally different privacy and architecture bet — the agent, its sandboxed compute, and its output-sharing mechanism all live entirely within the browser and direct peer connections.

Self-Hosting

Licensing Model Apache-2.0 licensed — fully open source with no license key.

Self-Hosting Restrictions Not applicable; there’s no backend to self-host. The extension runs entirely client-side with BYOK model access.

License Key Required No — you supply your own model provider API key (BYOK), not a license key for the extension itself.