Oryn

The browser built for AI agents.

---

What is Oryn?

Oryn (Open Runtime for Intentful Navigation) is a browser automation system designed specifically for AI agents. Instead of forcing agents to understand screenshots, parse HTML, or construct complex function calls, Oryn provides a semantic intent language that speaks naturally to how agents think about web interaction.

Traditional agent-browser interfaces fail because they expose the wrong abstraction. Oryn fixes this by providing:

• Semantic observations instead of raw HTML or pixels
• Intent language instead of rigid function schemas
• Consistent behavior across all deployment environments

The Unified CLI

Oryn provides a single unified binary oryn that adapts to any environment:

# Headless Mode (Cloud/CI)
oryn headless

# Embedded Mode (IoT/Edge)
oryn embedded --driver-url http://localhost:8080

# Remote Mode (Debug/Assistance)
oryn remote --port 9001

The Oryn2 native runtime is under gated implementation. It reuses the existing OIL parser and evaluation assets while replacing the browser dependency with an Oryn-owned DOM, semantic loop, and raw-V8 host. See the implementation status for exactly what is available today.

# Current native proof: load a local fixture, then enter OIL interactively
oryn native --html test-harness/scenarios/spa/svelte-tasks.html

# Or execute an OIL file against the fixture
oryn --file path/to/script.oil native --html path/to/document.html

# Verify the production runtime fingerprint (native V8 is enabled by default)
oryn native --runtime-info

# Build the deliberately non-production no-V8 probe only when measuring it
cargo build --release -p oryn --no-default-features --target-dir target/no-v8-probe

See the User Guide for full usage instructions.

Quick Example

Instead of parsing HTML or analyzing screenshots, agents interact naturally:

goto github.com/login
observe

@ github.com/login "Sign in to GitHub"
[1] input/email "Username or email" {required}
[2] input/password "Password" {required}  
[3] button/submit "Sign in" {primary}

type 1 "myusername"
type 2 "mypassword"
click 3

The agent sees labeled interactive elements and issues simple commands. No CSS selectors, no XPath, no DOM traversal—just intent.

Why Oryn?

The Problem

Current approaches force agents into unnatural workflows:

Screenshot/Vision: Expensive inference, unreliable text extraction, no understanding of interactive state

HTML Parsing: Thousands of tokens of markup, complex reasoning about visibility and interactivity

Function Calls: Rigid schemas, verbose definitions, no tolerance for natural variation

The Solution

Oryn presents the web as agents need to see it:

• Interactive elements are labeled and typed
• State is explicit (required, disabled, checked)
• Patterns are detected (login forms, search boxes)
• Commands are forgiving and semantic

Mode Selection Guide

Choose oryn-e (Embedded) when:

• Running on Raspberry Pi or IoT hardware
• Deploying to resource-constrained containers
• RAM is precious (~50MB footprint)
• WebKit compatibility is sufficient

Choose oryn-h (Headless) when:

• Running cloud-based automation
• Maximum browser compatibility needed (~99%)
• Handling complex SPAs
• Network interception required

Choose oryn-r (Remote) when:

• Agent needs user's logged-in sessions
• Anti-bot bypass is critical (real browser fingerprint)
• User wants to watch agent actions
• Interactive assistance workflows

Choose oryn-w (WASM Extension) when:

• No server/backend infrastructure available
• Instant setup required (load extension, go)
• Quick one-off commands on any webpage
• Minimal resource footprint needed (<10MB memory)
• See extension-w/README.md for details

Documentation

Document	Description
USER_GUIDE.md	Full installation and usage guide
GOOGLE_ADK_TUTORIAL.md	Integration with Google ADK Agents
SPEC-INTENT-LANGUAGE.md	The agent-facing command protocol
SPEC-SCANNER-PROTOCOL.md	Internal browser-scanner interface
SPEC-UNIFIED.md	Architecture and mode comparison
PRODUCT-INTRO.md	Product overview and vision

CI and Release Workflows

GitHub Actions workflows are defined under .github/workflows/:

• ci-rust.yml: cargo fmt, cargo clippy, cargo test on PRs and main.
• ci-js.yml: scanner checks plus extension-w lint/tests on PRs and main.
• ci-e2e-quick.yml: quick Docker-backed E2E smoke suite on PRs and main.
• nightly-full-e2e.yml: scheduled/manual full E2E matrix (oryn-h, oryn-e-*, oryn-r).
• preview-release.yml: tag-driven preview packaging and release publishing for preview-v* tags.
• docs-pages.yml: MkDocs build and deploy to GitHub Pages.

For local workflow validation with act, repo defaults are in .actrc.

Project Structure

oryn/
├── crates/
│   ├── oryn-core/          # Shared protocol and types
│   ├── oryn-scanner/       # Universal JavaScript scanner
│   ├── oryn-native/        # Oryn2 native runtime proof
│   ├── oryn-webidl/        # Oryn2 WebIDL parser/generator spine
│   ├── oryn-e/             # Embedded mode binary
│   ├── oryn-h/             # Headless mode binary
│   └── oryn-r/             # Remote mode binary
├── extension/                 # Browser extension for remote mode
├── docs/
│   ├── SPEC-INTENT-LANGUAGE.md
│   ├── SPEC-SCANNER-PROTOCOL.md
│   └── SPEC-UNIFIED.md
└── README.md

Core Concepts

Universal Scanner

A single JavaScript implementation runs inside the compatibility browser contexts—WebKit, Chromium, and browser extensions. Oryn2 native mode instead parses HTML into its Rust-owned DOM and produces v2 semantic observations directly; it does not duplicate or inject the compatibility scanner.

Intent Language

Commands are designed for agent ergonomics:

• Case-insensitive, forgiving syntax
• Multiple targeting strategies (ID, text, role, selector)
• Multi-level abstraction (direct → semantic → intent)
• Token-efficient responses

Backend Trait

All binaries implement the same interface. Agents can switch modes without changing their logic. The same scanner code, the same protocol, the same behavior.

Requirements

oryn-e (Embedded)

• Linux with WPE WebKit / COG
• ~50MB RAM available

oryn-h (Headless)

• Chromium browser installed
• ~300MB+ RAM available

oryn-r (Remote)

• Browser extension installed
• WebSocket connection to server

License

Existing Oryn code is available under the MIT License. New Oryn2-owned components may opt into MIT OR Apache-2.0; those components declare that choice in their package metadata.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

---

Oryn: Intent, not implementation.