neuron-core/neuron-ai

Neuron AI – PHP framework for building agentic applications

What it is

• A Composer‑installable PHP library that lets you create, orchestrate and run AI agents directly inside a PHP codebase.
• Provides first‑class abstractions for LLM providers, memory, toolkits, RAG, structured output, multi‑agent workflows and human‑in‑the‑loop handling.
• Targets developers who want the agent to be the core architecture of their app rather than an after‑thought feature.

Key concepts & components

Component	Role
Agent (extends NeuronAI\Agent\Agent)	Base class for a conversational AI with built‑in memory, streaming and tool calling.
RAG (extends NeuronAI\RAG\RAG)	Combines a provider, an embeddings service and a vector store (e.g., Pinecone) for retrieval‑augmented generation.
Tools / Toolkits	PHP objects that expose concrete actions (e.g., MySQL queries) to the LLM via function‑calling.
MCP Connector	Connects to an external Model‑Context‑Protocol server so tools can be served remotely.
Workflow	A low‑level, component‑based pipeline that lets you wire providers, loaders, vector stores, etc., into custom agentic flows with checkpointing and human‑in‑the‑loop steps.
Structured Output	Decorates a PHP class with #[SchemaProperty] attributes; the agent returns a typed object instead of raw text.

Supported LLM providers (plug‑and‑play via a single provider() method): Anthropic, OpenAI (including Azure and compatible APIs), Ollama, Gemini, Mistral, HuggingFace, Deepseek, Grok, AWS Bedrock, Cohere, ZAI, Alibaba DashScope, plus a generic Neuron Router.

Typical workflow

1. Install – composer require neuron-core/neuron-ai.
2. Add skills – npx skills add ./vendor/neuron-core/neuron-ai/skills -y (symlinks keep them up‑to‑date).
3. Pick framework integration – run the appropriate skill (neuron-laravel-integration or neuron-symfony-integration) or use the plain PHP skill.
4. Create an agent – vendor/bin/neuron make:agent MyAgent and implement provider(), instructions() and optionally tools().
5. Interact – instantiate the agent, set a thread ID, and call $agent->chat(new UserMessage('…')).
6. Extend – add toolkits, RAG, structured output or compose a full Workflow for complex, multi‑step processes.

Example – a data‑analyst agent

class DataAnalystAgent extends Agent {
    protected function provider(): AIProviderInterface {
        return new Anthropic(key: 'ANTHROPIC_API_KEY', model: 'ANTHROPIC_MODEL');
    }
    protected function instructions(): SystemMessage {
        return new SystemMessage('You are a data analyst expert in creating reports from SQL databases.');
    }
    protected function tools(): array {
        return [MySQLToolkit::make(\DB::connection()->getPdo())
                ->only([MySQLSchemaTool::class, MySQLSelectTool::class])];
    }
}

$resp = DataAnalystAgent::make()->setThreadId('demo')
        ->chat(new UserMessage('How many orders did we receive today?'))
        ->getMessage();
echo $resp->getContent();

The agent remembers the conversation, can safely run read‑only SQL, and can be swapped to another LLM provider with a single line change.

Ecosystem & community

• A registry of extensions (skills, integrations, toolkits) built specifically for agentic PHP apps.
• Official channels: docs site, newsletter, Inspector monitoring service, an e‑book, and social media.
• Designed as a vertical stack: the same Workflow that powers a single‑agent demo also scales to multi‑agent production systems.

Development & testing

• Requires PHP ≥ 8.1.
• Test suite runs in Docker Compose (docker compose up -d then vendor/bin/phpunit).
• Security issues are reported to support@inspector.dev.

Who should consider Neuron AI?

• PHP teams (Laravel, Symfony, or plain PHP) that want AI agents to be first‑class citizens of their architecture.
• Companies building internal tools, data‑analysis bots, or any product where the AI needs memory, tool access, or retrieval‑augmented generation.
• Developers looking for a unified way to switch LLM providers, add custom tools, and orchestrate complex workflows without stitching together disparate libraries.

---

All details are taken directly from the repository’s README; no additional features are inferred.