peerd – AI Agent Harness That Runs in Your Browser

TL;DR

TL;DR: peerd is a browser extension for Chrome and Firefox that runs a full AI agent loop inside the browser itself — sandboxed compute, BYOK model routing, no backend, no telemetry.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: peerd.ai ← verified
• Source repository: github.com/NotASithLord/peerd ← README read end-to-end
• License: Apache 2.0 ← verified via repo
• HN launch thread: news.ycombinator.com/item?id=48646165 (2026-06-23)
• Source last checked: 2026-07-07 (commit main branch)

What Is peerd?

peerd is a Chrome and Firefox browser extension that runs a complete AI agent loop inside the browser, with no backend and no telemetry in the data path.

Key characteristics (verbatim from README):

peerd is the first AI agent harness native to the browser. It reads and drives your pages, spins up sandboxed compute (JS Notebooks, compiled WebAssembly tools, full Linux VMs, personal client-side apps), and shares what it builds over a peer-to-peer WebRTC network built for agent-to-agent communication.

It is BYOK — bring your own keys to the model provider of your choice. No account, no cloud component.

Security Architecture

The README emphasises that security is built on the browser’s own sandbox model rather than custom cryptographic code:

• V8 isolates for sandboxing each actor
• WebCrypto for the vault, WebAuthn passkeys to unlock it
• Opaque-origin iframes for isolation between content and agent
• Subresource Integrity for tool loading
• Each browser tab, VM, notebook, and app is driven by its own actor — a separate agent loop that holds no key and only that one environment’s tools
• The main agent acts as orchestrator, delegating goals and receiving fenced summaries; raw page text and command output never reach the context holding your keys

Setup

Install from source (developer preview)

# Clone the repo
git clone https://github.com/NotASithLord/peerd.git
cd peerd

# Load the extension unpacked in Chrome
# 1. Go to chrome://extensions/
# 2. Enable "Developer mode" (top right)
# 3. Click "Load unpacked"
# 4. Select the extension/peerd-chrome directory

# For Firefox: use extension/peerd-firefox or about:debugging#/runtime/this-firefox

The extension requires no build step — vanilla JavaScript with Manifest V3.

How It Works

Agent loop

1. Per-tab actors — each browser tab runs its own isolated agent loop with its own toolset
2. Main orchestrator — the service worker holds your API keys and gates all requests
3. Observed effect verification — page actions report back what actually changed on the live page, so success is judged from observed effect rather than model assumption
4. P2P WebRTC (preview channel) — agents share what they build over a peer-to-peer network

Sandboxed compute options

From the README, the extension can spin up:

• JS Notebooks — inline script execution
• Compiled WebAssembly tools — native-speed sandboxed compute
• Full Linux VMs — heavier isolation when needed
• Personal client-side apps

Practical Evaluation Checklist

• [ ] Extension loads in Chrome and Firefox without errors
• [ ] API key is stored in the browser vault (not sent to any server)
• [ ] A task delegated to a tab actor returns an observed-effect summary
• [ ] Multiple tabs can run separate actors simultaneously
• [ ] The P2P sharing channel connects two browser instances
• [ ] Service worker correctly gates outbound requests
• [ ] Passkey unlocks the vault without password re-entry

Security Notes

• No backend in the data path — all agent computation runs client-side
• No telemetry — nothing is sent to peerd’s servers
• BYOK model routing — API keys stay in the browser, only requests go to your chosen provider
• Actor isolation — a confused or prompt-injected main agent has no direct tool to touch any environment; actors communicate only through the service worker gate
• Status: 0.x experimental — breaking changes are likely before v1.0

FAQ

Q: Does peerd work with any model provider? A: Yes — it is BYOK (bring your own keys). You wire in your own API keys for any provider that exposes a compatible API endpoint.

Q: Is there a hosted or cloud version? A: No. The entire agent loop runs in the browser. There is no cloud component in the data path.

Q: How does it compare to running a local agent via CLI? A: peerd runs entirely inside the browser tab you already use, with direct access to the live DOM and your existing sessions. CLI agents typically run headless and must re-implement page interaction from scratch.

Q: What happens if the service worker is killed by the browser? A: The README notes this is an open engineering challenge — service worker lifetime is bounded by the browser, so long-running agents may need to checkpoint state.

Q: Is this production-ready? A: No — status is 0.x experimental. Breaking changes are expected, and the developers make no v1 commitment yet.

Conclusion

peerd is a novel approach to AI agent runtimes: instead of running a backend agent that reaches out to a browser, it inverts the model and runs the agent inside the browser itself. The security architecture is grounded in the browser’s own primitives (V8 isolates, WebCrypto, WebAuthn), which is more credible than most “local-first” agents that still phone home for licensing. The 0.x status and “no v1 commitment” caveat mean it is firmly in the experimental category — useful for exploring agent-browser co-design, not ready to slot into a production workflow.