NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
Packagist, #1798 most downloaded on Packagist
A unified layer for building AI-native applications in PHP.
Latest stable
4.1.22
Last release 16 days ago
05 Nov 2026
Ships on a steady schedule
a new release about every 1 weeks
Nearly every release is documented
notes for 55 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
2 years old
486 releases · first in 2025
Drag across the ruler to compare, or tap a release to open it.
Fix chat hisotry when start woth tool messages: PR #386
Fix chat hisotry when start woth tool messages: PR #386
Nothing published for this version
Support deserialization for Structured Output objects with promoted properties. PR#384
Support deserialization for Structured Output objects with promoted properties. PR#384
Strenght the mysql schema tool description
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Nothing published for this version
Fix AssistantMessage mapping in Gemini
Fix AssistantMessage mapping in Gemini
return last message as generator result when streaming in RAG.
return last message as generator result when streaming in RAG.
Nothing published for this version
Fix tool max tries property name #361
Fix tool max tries property name #361
Nothing published for this version
This release introduce a new component to handle the agent chat history with eloquent model.
This release introduce a new component to handle the agent chat history with eloquent model.
https://docs.neuron-ai.dev/the-basics/chat-history-and-memory#eloquentchathisotry
As usual feel free to provide your feedback opening issue or publishing a PR.
It's also available a new persistence layer for Workflow, based on SQL database:
https://docs.neuron-ai.dev/workflow/persistence#databasepersistence
This allows cetrilized management of workflow persistence.
Catch streaming contents before tool calls - fix #353
Catch streaming contents before tool calls - fix #353
Nothing published for this version
Meilisearch vector store index auto-creation
This release brings the possibility to inject Guzzle middleware in providers class. It's useful to inspect what happens during LLM interactions. #347
This release brings the possibility to inject Guzzle middleware in providers class. It's useful to inspect what happens during LLM interactions. #347
Support nullable dimension in OpenAI embeddings provider #346
Support nullable dimension in OpenAI embeddings provider #346
Gracefully handle file chat history flush #337
Nothing published for this version
Nothing published for this version
New AWS Bedrock embeddings provider
New AWS Bedrock embeddings provider
Check out the documentation.
Update ChromaDB to v2 and introducing its unit tests.
Update ChromaDB to v2 and introducing its unit tests.
Nothing published for this version
Nothing published for this version
You can now run multiple RAG requests in parallel as it was already possible with the simple agent class.
You can now run multiple RAG requests in parallel as it was already possible with the simple agent class.
// Create separate agent instances
$rag1 = MyRAG::make();
$rag2 = MyRAG::make();
$rag3 = MyRAG::make();
// Execute multiple parallel requests
$results = Utils::unwrap([
'response_1' => $rag1->chatAsync(new UserMessage("Classify: Red cotton shirt, size M")),
'response_2' => $rag2->chatAsync(new UserMessage("Classify: wireless headphones, Bluetooth 5.3")),
'response_3' => $rag3->chatAsync(new UserMessage("Classify: laptop, Intel i7, 16GB RAM"))
]);
// Access results
echo $results['response_1']->getContent();
echo $results['response_2']->getContent();
echo $results['response_3']->getContent();
If your agents are tool-hungry, you can enable parallel execution if the model ask for multiple tool calls in a single request.
If your agents are tool-hungry, you can enable parallel execution if the model ask for multiple tool calls in a single request.
Just attach the ParallelToolCalls trait to your Agent:
class DemoAgent extends Agent
{
use ParallelToolCalls;
protected function provider(): AIProviderInterface
{
return new OpenAIResponses(
$_ENV['OPENAI_API_KEY'],
'gpt-4o-mini',
);
}
protected function tools(): array
{
return [
CalculatorToolkit::make(),
];
}
}
To use this trait you need to install the spatie/fork package. For more information check out the GitHub repository: https://github.com/spatie/fork
This implementation requires the pcntl extension which is installed in many Unix and Mac systems by default.
❗️ pcntl only works in CLI processes, not in a web context.
PR #319 enable the override of the tool calls execution.
PR #319 enable the override of the tool calls execution.
Nothing published for this version
Fix custom state serialization on workflow interruption
Fix custom state serialization on workflow interruption
Nothing published for this version
SSE (Server-Sent Events) is a mechanism that allows web clients to receive automatic updates from a server. Those updates are known as "events", and a
SSE (Server-Sent Events) is a mechanism that allows web clients to receive automatic updates from a server. Those updates are known as "events", and are sent over a single, long-lived HTTP connection.
To use the SSE transport you need to set async ⇒ true in the configuration parameters.
use NeuronAI\MCP\McpConnector;
class MyAgent extends Agent
{
...
protected function tools(): array
{
return [
...McpConnector::make([
'url' => 'https://mcp.example.com',
'token' => 'BEARER_TOKEN',
'timeout' => 30,
'async' => true
])->tools(),
];
}
}
Check out the official documentation.
Fix OpenAI responses API tool calls for tools with no arguments
Nothing published for this version
Fix array with multiple types validation
Fix array with multiple types validation
Support for arrays with multiple object types.
Support for arrays with multiple object types.
You have two options to specify the PHP classes you want to build the array with:
Using the square brackets syntax:
class Person
{
...
/**
* @var \App\TextBlock[]|\App\TableBlock[]
*/
#[SchemaProperty(description: 'The list of tag for the user profile.', required: true)]
#[ArrayOf([TextBlock::class, TableBlock::class])]
public array $tags;
}
Or using the "array<...>" syntax:
class Person
{
...
/**
* @var array<\App\TextBlock|\App\TableBlock>
*/
#[SchemaProperty(description: 'The list of tag for the user profile.', required: true)]
#[ArrayOf([TextBlock::class, TableBlock::class])]
public array $tags;
}
As you can notice from the examples above you can pass an array of class-string to the ArrayOf validation rule to ensure the final array will contains only instances of the listed classes.
fix openai responses API image/file attachment
fix openai responses API image/file attachment
Staring from #310 by @viicslen
Simplify human-in-the-loop pattern in Workflow:
Simplify human-in-the-loop pattern in Workflow:
https://docs.neuron-ai.dev/workflow/human-in-the-loop#checkpointing
Fix #309 - Unnecessary call to addToChatHistory in Structured APIs
Fix #309 - Unnecessary call to addToChatHistory in Structured APIs
pageCount() method on PdfReader class
Fix McpConnector bug if the tool provided by the server has no properties #306 by @RauchenwaldC
The tools inside the database related toolkits now ask the model to provide a parametrized query, and a list of parameters so the query can be properl
The tools inside the database related toolkits now ask the model to provide a parametrized query, and a list of parameters so the query can be properly prepared to avoid sql injection.
The constructor has a nullable class property but the class_exists check do not accept null raising an exception. The property now check if the class property is properly filled before the calling class_exists.
Fix #300
PR #303 from @zliebersbach
The tool inside the database related toolkits now ask the model to provide a parametrized query, and a listo of parameters so the query can be properl
The tool inside the database related toolkits now ask the model to provide a parametrized query, and a listo of parameters so the query can be properly prepared to avoid sql injection.
Nothing published for this version
Nothing published for this version
This release supports the ability to use custom workflow state class, so you can work with strictly typed properties.
This release supports the ability to use custom workflow state class, so you can work with strictly typed properties.
Create your state class extending the default WorkflowState:
use NeuronAI\Workflow\WorkflowState;
class CustomState extends WorkflowState
{
public string $custom = 'custom property';
}
Use this custom state as input type of the nodes __invoke mthod:
class MyNode extends Node
{
public function __invoke(StartEvent $event, CustomState $state): StopEvent
{
echo $state->custom;
return new StopEvent();
}
}
Finally start the Workflow passing the custom state instance. It will be propagated through the nodes:
$workflow = Workflow::make(new CustomState())->addNode($node);
$state = $workflow->start()->getResult();
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
OpenAI marked completions API as deprecated and now it's promoting the transition to responses API in order to take advantage of the latest features:…
OpenAI marked completions API as deprecated and now it's promoting the transition to responses API in order to take advantage of the latest features: https://platform.openai.com/docs/guides/migrate-to-responses
To keep the compatibility with all other provders that use the OpenAI completion API format, we implemented the Responses API a seprarate provider OpenAIResponses.
You can attach this provider to your agent as usual:
class MyAgent extends Agent
{
public function provider(): AIProviderInterface
{
return new OpenAIResponses(
key: 'OPENAI_API_KEY',
model: 'OPENAI_MODEL',
);
}
}
With the OpenAI responses API compatibility we decided to bring into the framework also the possibility to use provider specific tools like web_search, file_search, etc offered by many providers, like: OpenAI, Gemini, Anthropic.
To instruct your agent to use a specific provider tool you can add a ProviderTool class instance in the tools lists:
use NeuronAI\Tools\ProviderTool;
class MyAgent extends Agent
{
public function provider(): AIProviderInterface
{
return new OpenAIResponses(
key: 'OPENAI_API_KEY',
model: 'OPENAI_MODEL',
);
}
public function tools(): array
{
return [
ProviderTool:make(
type: 'web_search'
)->setOptions([...]),
// Add other tools and toolkits as usual...
];
}
}
Consider that providers has a lot of limitations in using their internal tools. The most flexible way to add capabilities to your agent remains the use of Tools & Toolkits.
Nothing published for this version
This release introduce the support to connect agents to remote MCP servers using URL instead of a commnand.
This release introduce the support to connect agents to remote MCP servers using URL instead of a commnand.
Remote servers are accessible via URLs and typically require authentication. You can use the token field in the configuration array, which will be used as an authorization token to communicate with the server:
class MyAgent extends Agent
{
protected function provider(): AIProviderInterface
{...}
protected function tools(): array
{
return [
...McpConnector::make([
'url' => 'https://mcp.example.com',
'token' => 'BEARER_TOKEN',
'timeout' => 10,
])->tools(),
];
}
}
The McpConnector automatically recognizes the URL and will use the StreamableHttpTransport to connect to the MCP server.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per month.