🦊 AgentFox

AgentFox is a high-velocity, persistent browser runtime built specifically for AI agents.

While traditional browsers are for humans, and frameworks like Playwright are for testers, AgentFox is for the agentic loop. It provides a daemon-backed, persistent browser surface that eliminates the latency of traditional automation tools.

---

🔥 Why AgentFox?

• Fast as Fuck: ~10ms command latency once the page is loaded. No cold starts, no overhead.
• Agent-Native Snapshots: afox view returns a Clean Markdown tree of interactive elements with stable IDs, pruned of utility noise (accessibility skips, cookie banners, etc.).
• Direct-ID Selection: High-speed interaction using [e1], [e2] style IDs that map directly to internal DOM elements, bypassing brittle CSS selectors.
• Persistent Session: Maintain cookies, logins, and JS state across multiple independent CLI calls.
• Lightweight: Powered by Rust and WebKitGTK. Meaningfully lower resource footprint than Chromium-based stacks.
• Zero Configuration: The daemon (afoxd) is automatically started by the CLI if it's not already running. It just works.

---

🚀 One-Line Installation

Install the AgentFox suite to ~/.local/bin:

curl -sSL https://raw.githubusercontent.com/MaenExists/AgentFox/main/install.sh | bash

Note: Requires Rust and WebKitGTK development headers (libwebkit2gtk-4.1-dev on Linux).

---

📖 The Agentic Loop

AgentFox is optimized for the Inspect -> Reason -> Act cycle.

1. Search & Navigate

No need to manually start a daemon. Just run a command.

afox search "latest news on autonomous agents"

2. Inspect (Agent-Friendly Markdown)

afox view

Returns a clean, machine-readable element tree:

# Google Search
URL: https://www.google.com/search?q=...

## [e10] Search Results
- [e15] (link) [Agents are the future](https://example.com/article)
- [e21] (input) Search: [ ... ]
- [e25] <button> Search </button>

3. Act Instantly

afox click e15
afox fill e21 "agentfox vs playwright"

---

🛠 Commands

Command	Usage	Description
search	afox search <q>	High-speed navigation (URL or search query).
open	afox open <url>	Navigate current session to a specific URL.
view	afox view	Get a clean Markdown snapshot of the page.
snap	afox snap	Get the raw semantic JSON tree.
click	afox click <id>	Direct-ID browser interaction.
fill	afox fill <id> <v>	Instant form input with event triggering.
text	afox text <id>	Extract clean text content from an element.
eval	afox eval <code>	Escape hatch for raw JS execution.
auth	afox auth <key> [url] [model]	Set up an LLM API for on-the-fly summaries.
ping	afox ping	Check if the daemon is alive (starts it if not).
quit	afox quit	Shutdown the background daemon.

---

🧠 LLM Summarization

AgentFox can extract the core content of any page and return an ultra-fast LLM summary instead of the full layout. This drastically reduces the context window usage for your agents.

Setup (One-time)

Configure AgentFox with an OpenAI-compatible API key. The auth command accepts the API key, the base URL of the API, and the model name.

Using OpenCode Zen (Free):

afox auth <YOUR_API_KEY> https://opencode.ai/zen/v1 nemotron-3-super-free

Using OpenAI:

afox auth sk-... https://api.openai.com/v1 gpt-4o-mini

Using Groq (Ultra-fast):

afox auth gsk_... https://api.groq.com/openai/v1 llama3-8b-8192

Using Local Ollama:

afox auth dummy_key http://localhost:11434/v1 llama3

You can provide any API key, endpoint, and model as long as the provider adheres to the standard OpenAI chat completions API format (/v1/chat/completions).

Usage

Append --summarize or -s to any navigation or snapshot command.

afox search "latest news on autonomous agents" --summarize
afox open https://example.com -s
afox view --summarize

🏗 Architecture

AgentFox uses a client-daemon architecture to eliminate the overhead of traditional browser control:

• cli/: The afox interface (Rust). Starts afoxd automatically.
• daemon/: The afoxd browser engine (Rust + WebKitGTK).
• protocol/: Shared high-speed JSON protocol for CLI-daemon IPC.

---

📜 License

AgentFox is open-source software licensed under the MIT License.