The first web-native AI agent harness

peerd is the first general-purpose agent runtime built directly on browser primitives: Workers, origins, sandboxing, OPFS, WASM/WASI, WebRTC, WebAuthn, and WebExtensions. It runs completely inside Chrome and Firefox, with your tabs, signed-in sessions, web apps, and local compute.

While agent platforms are trying to pull the browser into the harness. peerd pulls the harness into the browser.

For actual inference you can choose a supported hosted model provider, a local model using localhost, or check out preliminary support for local WebGPU models (we're keeping an eye on WebNN as well).

No peerd account, hosted browser, or tool-server connection is required. Current builds send no product telemetry to peerd.

Install · peerd.ai · Architecture · Security

Features

• Works in the browser you already use. The agent can read and drive your tabs, web apps, signed-in sessions, and page content.
• Builds reusable site clients. The web actor can learn a site once and use that client again on later tasks.
• Runs code inside browser boundaries. Scripts, sealed JavaScript Notebooks, compiled WASI tools, browser Apps, and Linux WebVMs give the agent local compute without access to your host operating system.
• Delegates to separate actors. Every page and compute environment gets its own keyless actor with tools scoped to that environment.
• Keeps useful context. Sessions, memory, skills, goals, review, and checkpoints live in the extension.
• Uses the model you choose. The live provider inventory is defined in registry.js, including BYOK cloud adapters and keyless local options.
• Connects browsers directly. Preview builds add signed identity, browser-to-browser discovery, dwapps, and agent-to-agent communication over WebRTC; store packages prune it entirely.

Why the browser

Local agents can access your whole computer. Remote agents live in someone else's. The browser is the alternative: local capability behind security boundaries hardened over three decades.

peerd uses those boundaries. Page work goes to separate actors with only the tools for that tab or environment. Credentials, network rules, confirmations, and audit stay with the extension. Its defense-in-depth design assumes unsafe content will eventually get past a filter.

Browser support

peerd supports Chromium and Firefox. Firefox runs actors in dedicated workers and uses visible Notebooks for JavaScript compute. Features that need Chrome's offscreen document host are removed from Firefox controls and model tools before use. Firefox preview builds omit dweb until Firefox has a mesh host.

Apps and WebVMs run on Chrome. Apps have no ambient network access. Remote resources, fetches, WebRTC, forms, and external document navigation are blocked. External HTTP and HTTPS links require user confirmation.

Concrete browser capability gaps, their upstream issues, and the tests required to remove each guard are tracked in docs/BROWSER-COMPATIBILITY.md.

The code is the source of truth for current behavior. Start with CLAUDE.md, then read the relevant module under extension/.

Security model

peerd uses browser isolation, narrow tool exposure, service-worker policy gates, and explicit egress controls. The main agent delegates environment work to keyless actors. On Chrome and Firefox, non-orchestrator agent loops run in separate dedicated worker heaps. If the browser cannot prove that boundary, the actor request does not run and performs no work on its target.

Network behavior depends on the operation. Model calls, web reads, runtime asset loads, sandbox traffic, and preview dweb traffic use different scoped paths and policies. See SECURITY.md and the threat model for the current boundaries and known limitations.

Install

Chrome from source

1. Clone the repository.
2. Open chrome://extensions.
3. Enable Developer mode.
4. Choose Load unpacked and select the extension/ directory.

Reload the extension from chrome://extensions after source changes.

Firefox from source

Firefox needs a Firefox-specific package. Do not load the checked-in Chrome development manifest. Use a Firefox version at or above the minimum declared in the channel patch under manifests/. That floor tracks the document-bound scripting support used by browser tools.

bun run package -- --channel=preview --browser=firefox --no-sign

Open about:debugging#/runtime/this-firefox, choose Load Temporary Add-on, and select artifacts/peerd-preview-firefox.xpi. Temporary add-ons must be loaded again after Firefox restarts. Browser and channel transforms are defined by the packaging scripts.

Release packages

See GitHub Releases for current artifacts. Store and preview builds differ. Store builds omit the dweb. Preview builds include it and may enable additional automation features. The packaging code is the authority for each browser and channel.

First run

1. Open peerd from the browser toolbar.
2. Create and unlock the local vault. Passphrase unlock is always available. Passkey unlock depends on WebAuthn PRF support in the browser and device.
3. Complete the short profile onboarding.
4. Open Settings, then add a provider key or choose a supported local provider.
5. Select a model and start a chat.

Only vault secrets and protected security records are covered by the vault encryption boundary. Other local extension state follows the storage rules in the security documentation.

Architecture

The extension has five main modules. Each module exposes its public API through its index.js.

The extension chassis lives in background/, offscreen/, sidepanel/, engine-tabs/, permissions/, shared/, and related support directories. Host placement and the cold-worker rule are documented in docs/EXTENSION-HOSTS.md.

Development

The source extension is vanilla JavaScript with ES modules and runs directly when loaded unpacked. There is no development bundler, transpiler, watcher, or generated runtime tree. Release packaging uses Bun only on its disposable staging copy to remove whitespace/comments from authored modules in the static service-worker and Chrome offscreen cold graphs. It preserves module boundaries, binding names, lazy imports, and every vendored byte. Pass --no-minify to bun run package -- ... when a readable diagnostic artifact is useful.

bun install
bun run gen:dev
bun test ./tests
bun scripts/cdp/run-inbrowser-tests.mjs
bun run typecheck
bun run lint
bun run e2e:verify
bun run preflight

There are three test surfaces:

• Bun tests for pure logic.
• In-browser tests for extension and browser integration. They run headless under Chrome and, sharded, under Gecko against the installed Firefox Store package. Each lane executes every test it registers; the totals differ slightly because some tests register only where a live service worker answers.
• Live Chrome E2E and visual verification for complete flows.

Alongside them, the red-team suite in tests/red-team/ drives each adversary from the threat model against the real defense code and records whether every hostile probe was blocked. Its matrix is docs/security/RED-TEAM-RESULTS.md.

Each lane publishes its own count as a badge above. The badge JSON under badges/ is generated by the CI job that ran the lane and then diffed, so a count on this page is always evidence from a run that happened rather than a number somebody typed. Regenerate one with bun run gen:badge:functional, bun run gen:badge:red-team, bun run gen:badge:inbrowser, bun run gen:badge:gecko (needs Firefox and geckodriver), or bun run gen:badge:e2e, and commit the result. bun run check:badges verifies the endpoints are well-formed without launching a browser.

For UI changes, run bun run e2e:verify, inspect scripts/cdp/artifacts/result.json, and inspect the generated screenshots.

Generated files must not be edited by hand. In particular, extension/manifest.json and extension/shared/channel-config.js come from the manifest and packaging sources. CI checks them for drift.

Read CONTRIBUTING.md before changing code.

Documentation

• CLAUDE.md: project structure, conventions, and current posture
• SECURITY.md: security policy and reporting
• docs/security/THREAT-MODEL.md: trust boundaries and residual risks
• docs/security/LIFECYCLE-CONTRACT.md: interruption behavior and recovery limits
• docs/security/RED-TEAM-RESULTS.md: red-team coverage
• docs/APP-ACTORS.md: manifest-defined App actors, live semantic adapters, and the in-tab co-pilot UX
• docs/DWAPP-BUNDLE.md: compressed dwapp transport, decoded working trees, and binary assets
• docs/store/: store packaging, permissions, privacy, and reviewer notes

Dependencies and license

The shipped extension has no npm runtime dependencies. package.json declares none, and packaging never resolves a node_modules path into the staged artifact, so the development-tool tree cannot reach an installed browser. Third-party runtime code is vendored under extension/vendor/ instead. Its source, version, and license live in the adjacent SOURCE.txt files, and every vendored byte is pinned by SHA-256 in extension/vendor/vendor.lock.json, which bun run check:vendor verifies in CI and preflight.

Two more supply-chain positions carry their own badges above, both regenerated by bun run gen:dev and drift-checked in CI. Every third-party GitHub Action runs at a full commit SHA, gated by check:actions: a major tag is a mutable ref its publisher can move, which would mean arbitrary code in a job holding this checkout, and in the release workflow, the signing secrets. Newly resolved dependencies also sit out a quarantine window before they can enter the lock, set by minimumReleaseAge in bunfig.toml, alongside a malware scan on install.

peerd is licensed under the Apache License 2.0. Vendored components retain their own licenses. CheerpX is a proprietary runtime provided by Leaning Technologies and is not covered by peerd's Apache license.