loki-auto MCP Server
Enables LLM agents to automate browser tasks through secure, sandboxed Rhai scripts executed in a WebAssembly VM, supporting DOM interactions and tab management via MCP tools.
README
loki-auto (Open-Source Edition)
loki-auto is an ultra-lightweight, zero-dependency "Serverless in Browser" web automation runtime designed for the AI Agent era. It executes episodic, transient LLM tool actions inside a secure Rhai VM (compiled to WebAssembly) sandbox container directly within the browser's foreground page context, completely bypassing heavy headless browser resource overhead.
🛠 Architecture
loki-auto uses a stateless, event-driven Oneshot CGI paradigm:
graph TD
A[LLM Agent / Orchestrator] <-->|MCP Protocol over Stdio| B[Axum MCP Server Host]
B <-->|WebSocket ws://127.0.0.1:10402| C[Browser Extension BG Script]
C <-->|chrome.runtime Messages| D[Active Tab Content Script]
D -->|Instantiates On-Demand| E[Rhai WASM VM Sandbox]
E -->|Synchronous DOM interactions| F[Page DOM]
- Layer 1: Host Environment – Rust Axum MCP server handling Model Context Protocol (MCP) tool bindings and bridging commands over local WebSockets.
- Layer 2: Browser Background – Extension background script managing tab states and routing target instructions.
- Layer 3: Browser Sandbox – The content script instantiates a lightweight WASM-compiled Rhai VM per prompt invocation, executes DOM operations, captures results/logs, and purges the VM from memory instantly.
🚀 Getting Started
1. Prerequisites
Ensure you have the following installed on your system:
- Bun (Workspace package manager)
- Rust & Cargo (with
wasm32-unknown-unknowntarget installed:rustup target add wasm32-unknown-unknown) - wasm-pack (for Rust-to-WASM compilation)
2. Compilation and Build
Build the WebAssembly sandbox first, and then compile the extension assets:
# 1. Compile Rust Sandbox to WebAssembly
cd packages/sandbox
wasm-pack build --target web
# 2. Return to root, install dependencies and compile the Chrome/Firefox extensions
cd ../..
bun install
bun run build
This will output two fully-compiled, self-contained extension directories under the root folder:
- Chrome (Manifest V3):
./dist/chrome - Firefox (Manifest V2):
./dist/firefox
3. Load the Browser Extension
- Firefox: Open
about:debugging#/runtime/this-firefox, click "Load Temporary Add-on...", and select the compiled packageloki-auto.xpiat the root of the project (or select./dist/firefox/manifest.json). - Chrome: Open
chrome://extensions/, enable Developer mode, click "Load unpacked", and select the./dist/chromefolder.
4. Launch the Local MCP Server
Start the Axum host to begin listening for browser connections and LLM tool calls:
# Start the MCP host (default port: 10402)
cargo run --bin loki-mcp-server
🔌 MCP Tool Specifications
loki-auto exposes standard Model Context Protocol (MCP) tools:
list_tabs- Lists all open browser tabs and their focused states.open_tab(url, active)- Opens a new tab with the specified URL.close_tab(tab_id)- Closes a target tab by ID.activate_tab(tab_id)- Focuses a browser tab to the foreground.execute_loki_oneshot(target_tab_id, target_url_pattern, rhai_script, payload)- Runs a Rhai automation script synchronously inside the sandboxed VM of the target tab.
📝 Rhai Sandbox DOM APIs
Scripts running inside the WASM sandbox have access to the following synchronous DOM hooks:
sleep(ms)- Suspends execution for the specified milliseconds.log(message)- Prints a message/object to the execution console logs.dom_to_string() -> String- Returns a cleaned, token-optimized (AEO-optimized) HTML tree of the current page.element_exists(selector) -> bool- Immediate check if a element matches the CSS selector.wait_element(selector, [timeout_ms]) -> bool- Blocks until the element is present, or throws a timeout exception.click(selector, [timeout_ms]) -> bool- Wait and click on the selected element.type_text(selector, text, [timeout_ms]) -> bool- Wait and instantly set text into an input or textarea, triggering input events (recommended for long texts/articles).type_as_human(selector, text, [timeout_ms]) -> bool- Wait and type text character-by-character with randomized human typing delays and standard keydown/keyup events (recommended for search boxes and login inputs).get_text(selector, [timeout_ms]) -> String- Extract inner text content.get_value(selector, [timeout_ms]) -> String- Extract form input value.get_attribute(selector, attr, [timeout_ms]) -> String- Extract specific attribute value.scroll_to(selector, [timeout_ms]) -> bool- Scroll the viewport smoothly to the targeted element.get_loki_data(dom_selector, [timeout_ms]) -> String- Scopes target DOM and extracts child nodes bearingdata-lokiattributes in a clean, Markdown-friendly format.
[!NOTE] The
[timeout_ms]parameter in DOM APIs (e.g.,click,type_text,type_as_human) specifies the maximum wait time for the target element to appear in the DOM before throwing aTimeoutError(defaults to5000ms). Fortype_as_human, this is the element detection limit and does not restrict the typing execution time itself (which runs asynchronously and is only limited by the global 45-second MCP oneshot task timeout).
Example Rhai Script
print("Interacting with search input...");
if wait_element("textarea[name='q']", 2000) {
type_text("textarea[name='q']", "weather in Boston today");
sleep(500);
click("input[name='btnK']");
}
🛡 Security and Isolation
- Strict Sandbox: The Rhai runtime inside WebAssembly has no access to standard JS APIs, cookies, localStorage, or system files unless explicitly exposed through bridge bindings.
- No Network Access: The sandbox cannot perform
fetchor AJAX requests. All actions occur inside the page context, and results are returned via the secure extension channel.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.