Create Full-Featured Agentic Applications in PHP

Before moving on, support the Neuron AI community giving a GitHub star ⭐️. Thank you!

Neuron is a PHP framework for creating and orchestrating AI Agents. It allows you to integrate AI entities in your PHP applications with a powerful and flexible architecture. We provide tools for the entire agentic application development lifecycle, from LLM interfaces, data loading, multi-agent orchestration, to monitoring and debugging. In addition, we provide tutorials and other educational content to help you get started using AI Agents in your projects.

Video Tutorial

• PHP: ^8.1

Go to the official documentation

Check out the technical guides and tutorials archive to learn how to start creating your AI Agents with Neuron https://docs.neuron-ai.dev/overview/fast-learning-by-video.

Neuron is the perfect AI architecture for your project.

Neuron offers a well-defined encapsulation pattern, allowing you to work on your agentic system in dedicated namespaces. You can enjoy the exact same experience of the other ecosystem packages you already love, like Filament, or Nova.

Example project (GitHub)

All Neuron components belong to its own interface, so you can define dependencies and automate objects creation using the Symfony service container. Watch how it works in a real project.

Symfony & Neuron (YouTube)

• Install
• Create an Agent
• Talk to the Agent
• Monitoring
• AI Providers
• Tools & Toolkits
• MCP Connector
• Structured Output
• RAG
• Workflow
• AI Assisted Development
• Security Vulnerabilities
• Official Documentation

Install the latest version via composer:

composer require neuron-core/neuron-ai

Neuron provides you with the Agent class you can extend to inherit the main features of the framework and create fully functional agents. This class automatically manages some advanced mechanisms for you, such as memory, tools, and function calls, up to RAG (Retrieval Augmented Generation). You can go deeper into these aspects in the documentation.

Let's create an Agent with the command below:

vendor/bin/neuron make:agent DataAnalystAgent

<?php

namespace App\Neuron;

use NeuronAI\Agent\Agent;
use NeuronAI\Agent\SystemPrompt;
use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\Anthropic\Anthropic;

class DataAnalystAgent extends Agent
{
    protected function provider(): AIProviderInterface
    {
        return new Anthropic(
            key: 'ANTHROPIC_API_KEY',
            model: 'ANTHROPIC_MODEL',
        );
    }

    protected function instructions(): string
    {
        return (string) new SystemPrompt(
            background: [
                "You are a data analyst expert in creating reports from SQL databases."
            ]
        );
    }
}

The SystemPrompt class is designed to take your base instructions and build a consistent prompt for the underlying model reducing the effort for prompt engineering.

Send a message to the agent to get a response from the underlying LLM:

$agent = DataAnalystAgent::make();


$response = $agent->chat(
    new UserMessage("Hi, I'm Valerio. Who are you?")
)->getMessage();
echo $response->getContent();
// I'm a data analyst. How can I help you today?


$response = $agent->chat(
    new UserMessage("Do you remember my name?")
)->getMessage();
echo $response->getContent();
// Your name is Valerio, as you said in your introduction.

As you can see in the example above, the Agent has memory of the ongoing conversation. Learn more about memory in the documentation.

Integrating AI Agents into your application, you’re not working only with functions and deterministic code, you program your agent influencing probability distributions. Same input ≠ output. That means reproducibility, versioning, and debugging become real problems.

Many of the Agents you build with Neuron will contain multiple steps with multiple invocations of LLM calls, tool usage, access to external memories, etc. As these applications get more and more complex, it becomes crucial to be able to inspect what exactly your agent is doing and why.

Why is the model taking certain decisions? What data is the model reacting to? Prompting is not programming in the common sense. No static types, small changes break output, long prompts cost latency, and no two models behave exactly the same with the same prompt.

The best way to take your AI application under control is with Inspector. After you sign up, make sure to set the INSPECTOR_INGESTION_KEY variable in the application environment file to start monitoring:

INSPECTOR_INGESTION_KEY=fwe45gtxxxxxxxxxxxxxxxxxxxxxxxxxxxx

After configuring the environment variable, you will see the agent execution timeline in your Inspector dashboard.

Learn more about Monitoring in the documentation.

With Neuron, you can switch between LLM providers with just one line of code, without any impact on your agent implementation. Supported providers:

• Anthropic (supports prompt caching)
• OpenAI (also as an embeddings provider)
• OpenAI Responses API
• OpenAI on Azure
• OpenAILike (OpenAI compatible APIs)
• Ollama (also as an embeddings provider)
• Gemini (also as an embeddings provider)
• Gemini Vertex
• Mistral
• HuggingFace
• Deepseek
• Grok
• AWS Bedrock Runtime
• Cohere
• ZAI

Make your agent able to perform concrete tasks, like reading from a database, by adding tools or toolkits (collections of tools).

<?php

namespace App\Neuron;

use NeuronAI\Agent\Agent;
use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\Anthropic\Anthropic;
use NeuronAI\Agent\SystemPrompt;
use NeuronAI\Tools\ToolProperty;
use NeuronAI\Tools\Tool;
use NeuronAI\Tools\Toolkits\MySQL\MySQLToolkit;

class DataAnalystAgent extends Agent
{
    protected function provider(): AIProviderInterface
    {
        return new Anthropic(
            key: 'ANTHROPIC_API_KEY',
            model: 'ANTHROPIC_MODEL',
        );
    }

    protected function instructions(): string
    {
        return (string) new SystemPrompt(
            background: [
                "You are a data analyst expert in creating reports from SQL databases."
            ]
        );
    }

    protected function tools(): array
    {
        return [
            MySQLToolkit::make(
                \DB::connection()->getPdo()
            ),
        ];
    }
}

Ask the agent something about your database:

$response = DataAnalystAgent::make()->chat(
    new UserMessage("How many orders we received today?")
)->getMessage();

echo $response->getContent();

Learn more about Tools in the documentation.

Instead of implementing tools manually, you can connect tools exposed by an MCP server with the McpConnector component:

<?php

namespace App\Neuron;

use NeuronAI\Agent\Agent;
use NeuronAI\MCP\McpConnector;
use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\Anthropic\Anthropic;
use NeuronAI\Tools\ToolProperty;
use NeuronAI\Tools\Tool;

class DataAnalystAgent extends Agent
{
    protected function provider(): AIProviderInterface
    {
        ...
    }

    protected function instructions(): string
    {
        ...
    }

    protected function tools(): array
    {
        return [
            // Connect to an MCP server
            ...McpConnector::make([
                'command' => 'npx',
                'args' => ['-y', '@modelcontextprotocol/server-everything'],
            ])->tools(),
        ];
    }
}

Learn more about MCP connector in the documentation.

There are scenarios where you need Agents to understand natural language, but output in a structured format, like business processes automation, data extraction, etc. to use the output with other downstream systems.

use App\Neuron\MyAgent;
use NeuronAI\Chat\Messages\UserMessage;
use NeuronAI\StructuredOutput\SchemaProperty;

/*
 * Define the output structure as a PHP class.
 */
class Person
{
    #[SchemaProperty(
        description: 'The user name',
        required: true
    )]
    public string $name;

    #[SchemaProperty(
        description: 'What the user love to eat'
    )]
    public string $preference;
}

/*
 * Talk to the agent requiring the structured output
 */
$person = MyAgent::make()->structured(
    new UserMessage("I'm John and I like pizza!"),
    Person::class
);

echo $person->name ' like '.$person->preference;
// John like pizza

Learn more about Structured Output on the documentation.

To create a RAG, you need to attach some additional components other than the AI provider, such as a vector store, and an embeddings provider.

Let's create a RAG with the command below:

vendor/bin/neuron make:rag MyChatBot

Here is an example of a RAG implementation:

<?php

namespace App\Neuron;

use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\Anthropic\Anthropic;
use NeuronAI\RAG\Embeddings\EmbeddingsProviderInterface;
use NeuronAI\RAG\Embeddings\VoyageEmbeddingProvider;
use NeuronAI\RAG\RAG;
use NeuronAI\RAG\VectorStore\PineconeVectorStore;
use NeuronAI\RAG\VectorStore\VectorStoreInterface;

class MyChatBot extends RAG
{
    protected function provider(): AIProviderInterface
    {
        return new Anthropic(
            key: 'ANTHROPIC_API_KEY',
            model: 'ANTHROPIC_MODEL',
        );
    }

    protected function embeddings(): EmbeddingsProviderInterface
    {
        return new VoyageEmbeddingProvider(
            key: 'VOYAGE_API_KEY',
            model: 'VOYAGE_MODEL'
        );
    }

    protected function vectorStore(): VectorStoreInterface
    {
        return new PineconeVectorStore(
            key: 'PINECONE_API_KEY',
            indexUrl: 'PINECONE_INDEX_URL'
        );
    }
}

Learn more about RAG in the documentation.

Think of a Workflow as a smart flowchart for your AI applications. The idea behind Workflow is to allow developers to use all the Neuron components like AI providers, embeddings, data loaders, chat history, vector store, etc., as standalone components to create totally customized agentic systems.

Agent and RAG classes represent a ready to use implementation of the most common patterns when it comes to retrieval use cases, or tool calls, structured output, etc. Workflow allows you to program your agentic system completely from scratch. Agent and RAG can be used inside a Workflow to complete tasks as any other component if you need their built-in capabilities.

Neuron Workflow supports a robust human-in-the-loop pattern, enabling human intervention at any point in an automated process. This is especially useful in large language model (LLM)-driven applications where model output may require validation, correction, or additional context to complete the task.

Learn more about Workflow on the documentation.

When working with AI coding assistants like Claude Code, Opencode, Cursor, or other similar tools, you have two options to give the AI deep context about Neuron components. This leads to more accurate code suggestions, better understanding of component APIs, and fewer hallucinations when generating Neuron code:

• Skills
• MCP server

If you discover a security vulnerability within Neuron, please send an e-mail to the Inspector team via support@inspector.dev. All security vulnerabilities will be promptly addressed.

Go to the official documentation