Larapanda

Language versions: English | Türkçe

Larapanda is a type-safe Lightpanda SDK for Laravel and plain PHP applications. It provides named instance profiles, runtime resolution (auto, cli, docker), and immutable method-scoped option objects for fetch, serve, and mcp.

Installation

Prerequisites

Larapanda uses a runtime resolver (auto, cli, docker). CLI execution (and auto mode when CLI is selected) requires a valid Lightpanda binary. Install Lightpanda from the official source: lightpanda.io.

curl -fsSL https://pkg.lightpanda.io/install.sh | bash

Install the Larapanda SDK package:

composer require ferdiunal/larapanda

For Laravel applications, publish the package configuration:

php artisan vendor:publish --tag="larapanda-config"

Optional integration dependencies:

composer require laravel/ai laravel/mcp

Configuration

Configuration is profile-based. Each client resolves an instance profile by name, and each profile may override the global defaults.

return [
    'default_instance' => 'default',

    'defaults' => [
        'runtime' => 'auto',
        'binary_path' => env('LARAPANDA_BINARY_PATH'),
        'docker' => [
            'command' => 'docker',
            'image' => 'lightpanda/browser:nightly',
            'container_name' => 'larapanda-lightpanda',
            'remove' => true,
            'extra_args' => [],
        ],
    ],

    'instances' => [
        'default' => [],
        'crawler' => [
            // For strict CLI mode:
            // 'runtime' => 'cli',
            // 'binary_path' => '/absolute/path/to/lightpanda',
        ],
        'mcp' => [
            // For strict Docker mode:
            // 'runtime' => 'docker',
        ],
    ],
];

• default_instance selects the instance profile used when no explicit profile name is provided.
• defaults defines baseline runtime settings shared by all instances unless overridden.
• instances.<name> contains per-profile overrides for runtime mode, binary path, Docker parameters, and process options.
• In auto runtime mode, Larapanda prefers CLI execution when binary_path is available and executable; otherwise it falls back to Docker.
• integrations.ai configures AI SDK tool adapters (instance, tool_prefix, exposed_tools, session settings).
• integrations.mcp configures Laravel MCP server tool adapters (instance, exposed_tools, session settings).

Usage

Scenario 1: Quickstart Fetch

Resolve the manager, select an instance profile, and execute a markdown fetch.

use Ferdiunal\Larapanda\Contracts\LarapandaManagerInterface;
use Ferdiunal\Larapanda\Enums\FetchDumpFormat;

$manager = app(LarapandaManagerInterface::class);
$client = $manager->instance('default');

$fetch = $client->fetchRequest('https://example.com')
    ->withOptions(
        dump: FetchDumpFormat::Markdown,
        obeyRobots: true,
        waitMs: 2000,
    )
    ->run();

$markdown = $fetch->asMarkdown();

Scenario 2: Output Modes

Use strict typed accessors based on the selected dump format.

use Ferdiunal\Larapanda\Enums\FetchDumpFormat;

$markdownResult = $client->fetchRequest('https://example.com')
    ->withOptions(
        dump: FetchDumpFormat::Markdown,
    )
    ->run();

$semanticTreeResult = $client->fetchRequest('https://example.com')
    ->withOptions(
        dump: FetchDumpFormat::SemanticTree,
    )
    ->run();

$semanticTreeTextResult = $client->fetchRequest('https://example.com')
    ->withOptions(
        dump: FetchDumpFormat::SemanticTreeText,
    )
    ->run();

$markdown = $markdownResult->asMarkdown();
$semanticTree = $semanticTreeResult->asSemanticTree();        // array<string, mixed>
$semanticTreeText = $semanticTreeTextResult->asSemanticTreeText();
$rawOutput = $semanticTreeTextResult->output();               // raw stdout fallback

FetchResult throws UnexpectedFetchOutputFormatException when a strict accessor does not match the selected dump format.

Scenario 3: Named Instance Profiles

Select the profile according to runtime intent:

• default: baseline fetch workloads.
• crawler: stricter crawl profile (for example, dedicated CLI runtime settings).
• mcp: MCP-oriented profile for long-running interactive sessions.

$defaultClient = $manager->instance('default');
$crawlerClient = $manager->instance('crawler');
$mcpClient = $manager->instance('mcp');

Scenario 4: Proxy-Aware Fetch

Use request-level proxy settings for one fetch operation:

use Ferdiunal\Larapanda\Enums\FetchDumpFormat;

$fetch = $client->fetchRequest('https://example.com')
    ->withOptions(
        dump: FetchDumpFormat::Markdown,
        httpProxy: 'http://127.0.0.1:3000',
        proxyBearerToken: 'MY-TOKEN',
    )
    ->run();

Use integration-level proxy settings for AI SDK and MCP server tool sessions:

// config/larapanda.php
'integrations' => [
    'ai' => [
        'http_proxy' => 'http://127.0.0.1:3000',
        'proxy_bearer_token' => 'MY-TOKEN',
    ],
    'mcp' => [
        'http_proxy' => 'http://127.0.0.1:3000',
        'proxy_bearer_token' => 'MY-TOKEN',
    ],
],

Scenario 5: Long-Running Serve and MCP Modes

Use RunningInstanceHandle lifecycle methods for process-safe execution:

$serveHandle = $client->serveRequest()
    ->withOptions(host: '127.0.0.1', port: 9222)
    ->run();

try {
    if ($serveHandle->isRunning()) {
        // connect with CDP client
    }
} finally {
    $serveHandle->stop();
    $serveHandle->wait(2.0);
}

$mcpHandle = $client->mcpRequest()->run();

try {
    if ($mcpHandle->isRunning()) {
        // attach MCP host/client
    }
} finally {
    $mcpHandle->stop();
    $mcpHandle->wait(2.0);
}

Common Patterns

// Robots-compliant fetch
$result = $client->fetchRequest('https://example.com')
    ->withOptions(obeyRobots: true, dump: FetchDumpFormat::Markdown)
    ->run();

// Strict accessor mismatch raises UnexpectedFetchOutputFormatException
$result->asSemanticTreeText();

Laravel AI SDK Tools

Install optional dependencies before using adapter features:

composer require laravel/ai laravel/mcp

Larapanda exposes AI SDK-compatible tools via LarapandaAiTools. The adapter is MCP-backed, session-aware, and config-driven.

Scenario 1: Full Tool Catalog

use Ferdiunal\Larapanda\Integrations\Ai\LarapandaAiTools;
use Illuminate\Support\Facades\AI;

$response = AI::provider('openai')
    ->model('gpt-5-mini')
    ->prompt('Open laravel.com and return the main headings.')
    ->tools(app(LarapandaAiTools::class)->make())
    ->text();

Tool naming uses the configured prefix (default lightpanda_), for example: lightpanda_markdown, lightpanda_semantic_tree, lightpanda_click.

Scenario 2: Restrict Exposed AI Tools

Limit model access to a subset of tool surfaces:

// config/larapanda.php
'integrations' => [
    'ai' => [
        'exposed_tools' => ['goto', 'markdown', 'semantic_tree'],
    ],
],

Scenario 3: Session-Aware AI Tool Continuity

For multi-step browsing tasks, guide the model to reuse a stable session_id across tool calls.

$response = AI::provider('openai')
    ->model('gpt-5-mini')
    ->prompt('Use lightpanda tools with session_id=\"docs-session\". First goto laravel.com, then return markdown.')
    ->tools(app(LarapandaAiTools::class)->make())
    ->text();

Laravel MCP Server (Optional / Advanced)

Use this layer when you want Laravel-managed MCP registration, container wiring, config-based tool filtering, and shared session/proxy policies.
If you only need a standalone MCP binary host, native lightpanda mcp can be used directly without Larapanda adapter classes.

Native vs Adapter Decision Matrix

Scenario 1: Register MCP Tools in routes/ai.php

use Ferdiunal\Larapanda\Integrations\Mcp\LarapandaMcpServer;

LarapandaMcpServer::registerLocal(name: 'lightpanda');

Scenario 2: Restrict MCP Tool Exposure

By default, all Lightpanda MCP tools are exposed. To reduce tool surface:

// config/larapanda.php
'integrations' => [
    'mcp' => [
        'exposed_tools' => ['goto', 'markdown', 'semantic_tree'],
    ],
],

Scenario 3: Session and Proxy Policy for MCP Adapter

Session and proxy behavior are controlled from config:

• Each integration path (AI and MCP server) maintains an isolated in-memory session pool.
• Sessions map to long-running lightpanda mcp processes.
• Sessions expire using session_ttl_seconds and are capped by max_sessions.
• session_id can be passed in tool arguments to preserve page context across calls.

// config/larapanda.php
'integrations' => [
    'mcp' => [
        'session_ttl_seconds' => 300,
        'max_sessions' => 32,
        'obey_robots' => true,
        'http_proxy' => 'http://127.0.0.1:3000',
        'proxy_bearer_token' => 'MY-TOKEN',
    ],
],

Scenario 4: Interactive MCP Argument Model (Native-Aligned)

Interactive tools use backend-node-driven arguments. The typical flow is:

1. Navigate and discover target nodes (goto + waitForSelector or interactiveElements).
2. Extract backendNodeId from the response payload.
3. Execute interaction tools with that node id on the same session_id.

// Example argument shapes (canonical tool contracts):
// goto|navigate:      ['url' => 'https://example.com', 'timeout' => 10000, 'waitUntil' => 'done']
// waitForSelector:    ['selector' => '#submit', 'timeout' => 5000]
// click|hover:        ['backendNodeId' => 123]
// fill:               ['backendNodeId' => 123, 'text' => 'Ferdi']
// press:              ['key' => 'Enter', 'backendNodeId' => 123] // backendNodeId optional
// selectOption:       ['backendNodeId' => 123, 'value' => 'tr']
// setChecked:         ['backendNodeId' => 123, 'checked' => true]
// scroll:             ['y' => 400, 'backendNodeId' => 123] // backendNodeId optional

Testing

Run the full test suite:

composer test

Run opt-in live CLI + MCP smoke tests (native MCP and Larapanda bridge):

LIGHTPANDA_LIVE_TESTS=1 \
LIGHTPANDA_BINARY_PATH=/Users/ferdiunal/Web/larapanda/lightpanda \
php vendor/bin/pest --group=live

Run MCP-focused live smoke tests only:

LIGHTPANDA_LIVE_TESTS=1 \
LIGHTPANDA_BINARY_PATH=/Users/ferdiunal/Web/larapanda/lightpanda \
php vendor/bin/pest --filter=Mcp --group=live

Live suite policy:

• Live tests are opt-in and are not required in the default CI pipeline.
• In constrained environments, DNS/port limitations result in deterministic skip, not flaky fail.
• Real protocol and argument contract mismatches continue to fail.

Prerequisites for live tests:

• LIGHTPANDA_BINARY_PATH points to a valid executable Lightpanda binary.
• The environment permits internet access and local port binding.
• Optional proxy smoke test requires LIGHTPANDA_HTTP_PROXY (and optional LIGHTPANDA_PROXY_BEARER_TOKEN).

Optional proxy smoke run:

LIGHTPANDA_LIVE_TESTS=1 \
LIGHTPANDA_BINARY_PATH=/Users/ferdiunal/Web/larapanda/lightpanda \
LIGHTPANDA_HTTP_PROXY=http://127.0.0.1:3000 \
LIGHTPANDA_PROXY_BEARER_TOKEN=YOUR_TOKEN \
php vendor/bin/pest --group=live

Credits

• Ferdi ÜNAL

License

This package is licensed under the MIT License. See LICENSE.md for details.