peerd – Browser-Native AI Agent Harness

TL;DR

TL;DR: peerd is a Chrome/Firefox extension that runs a full AI agent loop entirely inside your browser — no backend, no telemetry, BYOK to Anthropic/OpenRouter/Ollama.

Source and Accuracy Notes

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

• Project page: peerd.ai ← MUST visit and verify
• Source repository: github.com/NotASithLord/peerd ← MUST read README
• License: Apache-2.0 ← verified via GitHub API license.spdx_id
• Show HN thread: news.ycombinator.com/item?id= (HN submission was the project URL itself, peerd.ai)
• Source last checked: 2026-06-26 (commit main, pushed 2026-06-26T10:41:13Z)

What Is peerd?

peerd describes itself as “the first AI agent harness native to the browser.” It is a Chrome/Firefox extension that runs a complete AI agent loop inside the browser — reading and driving pages, spinning up sandboxed compute environments, and sharing outputs peer-to-peer over WebRTC.

Instead of routing agent traffic through a remote server, peerd keeps all data on your machine. The browser itself becomes the runtime, leveraging decades of hardened browser security primitives (V8 isolates for sandboxing, WebCrypto for the vault, WebAuthn passkeys, opaque-origin iframes, Subresource Integrity) rather than writing custom cryptographic or process-isolation code.

The README puts it directly:

“peerd is the first AI agent harness native to the browser. It’s a Chrome/Firefox extension that runs a full agent loop inside the browser you already use, with your existing tabs and sessions.”

Key design points from the source:

• No backend, no telemetry — nothing leaves your machine except API calls you explicitly authorize
• BYOK — bring your own key from Anthropic, OpenRouter, or a local Ollama instance
• WASM Linux VMs — spins up sandboxed Linux environments compiled to WebAssembly inside the browser
• JS Notebooks — disposable compute sandboxes
• Peer-to-peer A2A over WebRTC — share what you build with other peerd instances (preview channel)
• Manifest V3 — supports both Chrome and Firefox

Setup Workflow

Step 1: Install the Extension (Developer Preview)

peerd is currently in 0.x experimental beta. There is no stable store listing yet. Developer preview is the primary install path.

Firefox (recommended for preview):

1. Download peerd-preview-firefox.xpi from the GitHub Releases page
2. Click the file in Firefox — it is AMO-signed, installs like any extension, and auto-updates

Chrome on macOS / Windows:

Chrome hard-disables off-store CRX installs on these platforms. The path is:

1. Download peerd-preview-chrome.zip from GitHub Releases
2. Unzip it to a folder on disk
3. Open chrome://extensions
4. Enable Developer mode (top-right toggle)
5. Click Load unpacked and select the unzipped folder

Note: No auto-update on this path. Download the new zip per release. The extension ID will be machine-specific.

Chrome on Linux (or policy-managed Chrome):

1. Download peerd-preview-chrome.crx
2. Enable Developer mode at chrome://extensions
3. Drag the .crx file onto the page — auto-update follows peerd.ai/updates/

Firefox Add-ons ID: [email protected] Chrome preview extension ID: lpdkhfeldihoejbbfonnbekpjclkknoc

Step 2: Configure Your Model Key

After installation:

1. Open the peerd extension popup or dashboard
2. Navigate to Settings → Vault (or equivalent)
3. Add your API key:
   • Anthropic key — for Claude models
   • OpenRouter key — for access to many models
   • Ollama — local inference, no key needed, nothing leaves your machine

Keys are encrypted locally in a browser vault and only ever sent to the provider you choose.

Step 3: Start an Agent Session

With the extension loaded and a key configured:

1. Click the peerd icon in your browser toolbar
2. The agent opens in a new tab or sidebar, reading your current tab context
3. Give it a task — it will drive your browser to complete it, step by step
4. Actions are verified against the live page before being marked complete

Step 4: Explore the Sandbox (Optional — Preview Channel)

The preview channel adds peer-to-peer sharing of builds over WebRTC:

1. Open Settings → Export & import to move state between store and preview installs
2. Other peerd peers can receive what you share over the dweb layer

Deeper Analysis

Security Model

peerd builds on browser platform primitives rather than custom crypto:

• V8 isolates provide sandboxing between the agent runner (no keys, no network) and the vault (holds your keys)
• WebCrypto encrypts the vault
• WebAuthn passkeys unlock the vault
• Opaque-origin iframes fence untrusted content
• Subresource Integrity validates external resources

The README states explicitly: “The agent that holds your keys never reads a raw page; a disposable runner with no keys and no network does, and its output comes back fenced as untrusted.”

No Build Step

peerd uses vanilla JavaScript — there is no build toolchain required. You load the extension/ folder directly into Chrome as unpacked source. This makes debugging and auditing straightforward.

Status: 0.x Experimental

The README is direct about this:

“Status: 0.x, experimental beta. Breaking changes are likely, storage formats may shift, and it drives your browser and holds your API keys, so use it with care.”

There is no V1 commitment. The surface will stabilize.

Practical Evaluation Checklist

• [ ] Extension loads in Chrome and Firefox without errors
• [ ] API key saved and encrypted in local vault
• [ ] Agent can read current browser tab context
• [ ] Agent can perform multi-step actions (click, type, navigate)
• [ ] WASM Linux VM sandbox executes without crashing the browser
• [ ] Ollama local inference works with no external network calls
• [ ] Data never leaves the browser except to the configured model provider
• [ ] Preview dweb channel connects peer-to-peer

Security Notes

• All data stays local — no telemetry, no backend, no cloud in the data path
• API keys are encrypted at rest using WebCrypto and only decrypted when sending a request to the configured provider
• The agent runner sandbox has no network access and no keys — it reads pages and returns structured output to the agent
• As a browser extension, peerd is subject to Chrome Web Store / AMO review processes for the store builds
• The developer preview (unpacked) has not been audited by a third party — review the source at github.com/NotASithLord/peerd before using with production keys

FAQ

Q: Does peerd work with any model provider? A: It supports Anthropic (Claude), OpenRouter (many models), and Ollama (local, no network). The BYOK model means you configure whichever provider you trust.

Q: Is there a stable version? A: No — peerd is 0.x experimental. The README warns that breaking changes are likely and storage formats may shift. Do not rely on it for production without careful evaluation.

Q: How is this different from a remote AI agent service? A: The agent loop runs inside your browser. No data passes through a peerd-owned server. You can use a local Ollama instance so API calls never leave your network at all.

Q: Does it work on mobile browsers? A: The current focus is desktop Chrome and Firefox. Mobile browser support is not documented.

Q: Can I self-host the update feed? A: The Chrome update feed at peerd.ai/updates/ is served by peerd. For full self-hosting isolation, use the unpacked load path and manually download new releases.

Conclusion

peerd is an interesting direction for browser-native AI agents — keeping the entire agent loop on the client, leveraging browser sandboxing rather than cloud VMs. The 0.x status means it is not production-ready, but the architecture is sound and the no-backend, no-telemetry promise is verifiable in the source.

For developers who want to experiment with AI agents that operate directly on their browser session — with full visibility and control over what data leaves their machine — peerd is worth watching.

Source and Accuracy Notes

• Project page: peerd.ai
• Source repository: github.com/NotASithLord/peerd
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• GitHub stars: 119 (as of 2026-06-26)
• Default branch: main, last pushed 2026-06-26