dartantic_ai

Dependencies

• Updated provider SDK constraints for Anthropic, Google AI, Mistral, Ollama, OpenAI, MCP, meta, uuid, and test.
• Updated generated schema examples to use json_serializable schema generation instead of soti_schema_plus.

Fixed

• Updated provider mappers for current SDK shapes, including Anthropic unknown content blocks, Mistral nullable/message-part content, Ollama typed keepAlive and stop values, OpenAI unknown finish reasons, OpenAI Responses file uploads, and Google/Ollama embeddings request shapes.

Fixed

• Google (Gemini) — server-side tools with client function tools — Requests that combine built-in server-side tools (Google Search, URL context, File Search, Maps grounding, code execution) with client-side function tools are rejected by Gemini unless tool_config.include_server_side_tool_invocations is true. GoogleChatModel now detects that combination and sets the flag on ToolConfig automatically.

Dependencies

• Google Generative AI SDK — Replaced google_cloud_ai_generativelanguage_v1beta with googleai_dart ^5.0.1 (#107). Chat, embeddings, and media generation now use this client; user-facing Dartantic APIs are unchanged aside from new options below.

Added

• Google (Gemini) — thinking levels — GoogleChatModelOptions.thinkingLevel (GoogleThinkingLevel: minimal, low, medium, high) for Gemini 3+ models that support thinking depth. Do not combine with thinkingBudgetTokens; the API rejects both. (#108)
• Google (Gemini) — File Search — GoogleChatModelOptions.fileSearch with GoogleFileSearchToolConfig (file search store names, optional topK and metadataFilter) for semantic retrieval from configured stores. (#108)
• Google (Gemini) — Maps grounding — GoogleChatModelOptions.mapsGrounding with GoogleMapsGroundingOptions (optional enableWidget for widget context in grounding metadata when supported). (#108)
• Google (Gemini) — grounding metadata — Model messages from Gemini can include grounding_metadata in message metadata (JSON from the API’s grounding metadata), including for Maps when enabled. (#108)
• xAI Responses — max_turns — XAIResponsesChatModelOptions.maxTurns maps to the API’s max_turns (cap on assistant / server-side tool iterations per request). See xAI’s tool documentation for interaction with client- vs server-side tools. (#106)

Dependencies

• dartantic_interface ^4.0.0 — adds ModelKind.video for video generation in model discovery. Exhaustive switches on ModelKind must handle video or use a default branch (see the dartantic_interface changelog).

Added

• Google (Gemini) — URL Context server-side tool — GoogleServerSideTool.urlContext in GoogleChatModelOptions.serverSideTools enables Gemini URL context so the model can retrieve and use content from allowed URLs.
• xAI (Grok) — xai provider (alias grok) using the OpenAI-compatible chat completions API at https://api.x.ai/v1, API key XAI_API_KEY. Chat, vision, tools, streaming, and typed output; embeddings and temperature are not supported in Dartantic for this provider.
• xAI Responses — xai-responses provider (alias grok-responses) using xAI’s Responses API with the same base URL and XAI_API_KEY. Supports thinking, server-side tools (web search, X search, file search, code interpreter, MCP), and native xAI image/video media generation with defaults grok-imagine-image / grok-imagine-video.

Documentation

• Provider and quick-start docs (including xAI), environment setup, README cross-links, and related wiki/docs site updates.

SDK Dependency Upgrades

Migrated all provider SDK dependencies to their latest major versions:

Breaking Changes

• Google Embeddings Model: Default embeddings model changed from text-embedding-004 to gemini-embedding-001. The old model was removed from Google's v1beta API.

Fixed: Google Server-Side Tools with Structured Output

Fixed issue #96 where combining Google server-side tools (Google Search, Code Execution) with typed output would fail with "Tool use with a response mime type: 'application/json' is unsupported".

Server-side tools now work with typed output using the same two-phase GoogleDoubleAgentOrchestrator approach as user-defined tools:

• Phase 1: Execute server-side tools (no outputSchema)
• Phase 2: Get structured JSON output (no tools)

// Now works! Previously failed with API error
final agent = Agent(
  'google',
  chatModelOptions: const GoogleChatModelOptions(
    serverSideTools: {GoogleServerSideTool.googleSearch},
  ),
);

final result = await agent.sendFor<MyOutput>(
  'Search for current weather and return as JSON',
  outputSchema: MyOutput.schema,
  outputFromJson: MyOutput.fromJson,
);

The fix automatically detects server-side tools and selects the appropriate orchestrator, with no code changes required for existing applications.

Streaming Thinking via chunk.thinking

Added a dedicated thinking field to ChatResult<String> for streaming thinking content. This provides symmetric access to thinking during streaming, matching how chunk.output provides streaming text:

await for (final chunk in agent.sendStream(prompt)) {
  if (chunk.thinking != null) {
    stdout.write(chunk.thinking);  // Real-time thinking display
  }
  stdout.write(chunk.output);  // Real-time text display
  history.addAll(chunk.messages);  // Consolidated messages
}

This is an additive change - the final consolidated message still contains ThinkingPart for history storage.

Breaking Change: Migrated to genai_primitives Types

The core message types have been migrated from custom implementations to the standardized genai_primitives package. This provides better interoperability with other GenAI tooling in the Dart ecosystem.

Types now re-exported from genai_primitives:

• ChatMessage, ChatMessageRole
• Part (alias for StandardPart), TextPart, DataPart, LinkPart, ThinkingPart
• ToolPart, ToolPartKind
• ToolDefinition

Note: Part is a typedef alias for StandardPart from genai_primitives 0.2.0. See dartantic_interface CHANGELOG for details on custom Part implementations.

Breaking Change: Migrated to json_schema_builder for Schemas

The Schema type is now provided by the json_schema_builder package instead of a custom implementation. This provides a more robust JSON Schema builder with better validation.

// NEW: Use S.object() for empty schemas, S.* for building schemas
import 'package:dartantic_ai/dartantic_ai.dart';

final tool = Tool(
  name: 'my_tool',
  description: 'Does something',
  inputSchema: S.object(properties: {
    'name': S.string(description: 'The name'),
  }),
  onCall: (args) => 'Hello ${args['name']}',
);

Thinking API

Extended thinking (chain-of-thought reasoning) is accessed via ChatResult.thinking for both streaming and non-streaming:

final agent = Agent('anthropic', enableThinking: true);

// Non-streaming
final result = await agent.send('Solve this puzzle...');
print(result.thinking);

// Streaming
await for (final chunk in agent.sendStream('Solve this puzzle...')) {
  if (chunk.thinking != null) stdout.write(chunk.thinking);  // Real-time
}

Thinking is also stored as ThinkingPart in message parts for conversation history.

Provider Changes

• Mistral Default Model: Changed default from mistral-small-latest to mistral-medium-latest for more reliable tool calling. The small model was truncating string arguments in certain scenarios.

Fixes

• Anthropic Thinking Metadata: The thinking signature is still stored in metadata while the thinking text is only stored in ThinkingPart.

• ThinkingPart Filtering: Each provider's message mapper now correctly handles ThinkingPart - Anthropic converts it to thinking blocks for the API, while other providers filter it out during mapping since they don't need thinking content sent back.

• Updated mcp_dart dependency to 1.2.1 to fix a null issue.

• Gemini 3 Tool Calling Fix: Fixed "Function call is missing a thought_signature" error when using tools with Gemini 3 models like gemini-3-flash-preview. No code changes required - just upgrade and it works. (#85). Thanks to @ElectricCookie for the PR!

• Mistral Tool Calling Support: Enhanced Mistral provider with robust tool calling capabilities:
  • Updated mistralai_dart dependency to 0.1.1+1 which fixes streaming tool call issues
  • Improved message mappers with null-safe tool call handling
  • Added filtering for incomplete tool calls in streaming responses
  • Enabled multiToolCalls capability for Mistral provider
  • Updated tests to verify tool calling works correctly with streaming
• Enhanced Cohere & Ollama Capabilities: Both providers now support multiToolCalls capability for parallel tool execution
• Added OCR (Optical Character Recognition) example to multimedia_input.dart demonstrating text extraction from images using Gemini's vision capabilities

• Added custom dimensions support for Mistral embeddings:
  • MistralEmbeddingsModel now passes outputDimension and encodingFormat parameters to the Mistral API
  • Leverages outputDimension parameter added in mistralai_dart PR #886
  • Updated tests to use codestral-embed-2505 for custom dimensions testing (default mistral-embed model doesn't support custom dimensions)
• Updated default Mistral chat model from open-mistral-7b to mistral-small-latest for better overall capabilities

• Added image editing support to media generation across all providers
  • Fixed GoogleMediaGenerationModel to accept attachments with image requests
  • Attachments (DataPart, LinkPart, TextPart) are now properly converted to Google API format (inlineData, fileData, text)
  • Enables image editing use cases: colorization, style transfer, inpainting
  • Added e2e tests for image editing with attachments for Google, OpenAI, and Anthropic providers
  • Updated all media generation examples to include image editing demonstrations
  • Refactored Google Part mapping to use shared mapPartsToGoogle() helper
• Refactored provider listModels() implementations to use SDK methods instead of raw HTTP:
  • Anthropic: Uses client.listModels() from anthropic_sdk_dart
  • Ollama: Uses client.listModels() from ollama_dart
  • Mistral: Uses client.listModels() from mistralai_dart
• Removed Anthropic signature_delta work-around (fixed in anthropic_sdk_dart 0.3.1)

• Updated Anthropic SDK compatibility for anthropic_sdk_dart 0.3.1:
  • ImageBlockSource now uses sealed class API with base64ImageSource() factory
  • Added support for new block types: DocumentBlock, RedactedThinkingBlock, ServerToolUseBlock, WebSearchToolResultBlock, MCPToolUseBlock
  • Added support for new delta types: SignatureBlockDelta, CitationsBlockDelta
  • Added pauseTurn and refusal stop reasons
• Updated Mistral SDK compatibility for mistralai_dart 0.1.1:
  • Fixed ambiguous imports for JsonSchema and Tool
  • Added error and toolCalls finish reasons

• updated dependencies

• updated dependencies

Breaking Change: Exposing dartantic_interface directly from dartantic_ai

It's no longer necessary to manually include the dartantic_interface package.

// OLD - had to import both packages
import 'package:dartantic_ai/dartantic_ai.dart';
import 'package:dartantic_interface/dartantic_interface.dart';

// NEW - one import does it all
import 'package:dartantic_ai/dartantic_ai.dart';

Breaking Change: Provider Factory Registry

Provider lookup has been moved from the Providers class to Agent static methods. Providers are now created via factory functions not cached instances.

// OLD
final provider = Providers.get('openai');
final allProviders = Providers.all;
Providers.providerMap['custom'] = MyProvider();
final provider2 = Providers.openai;

// NEW
final provider = Agent.getProvider('openai');
final allProviders = Agent.allProviders;
Agent.providerFactories['custom'] = MyProvider.new;
final provider2 = OpenAIProvider();

Breaking Change: Moved OpenAI-compat providers to example (except OpenRouter)

Removed the following intrinsic providers from dartantic to the openai_compat.dart example:

• google-openai
• together
• ollama-openai

The openrouter OpenAI-compatible provider remains as an intrinsic provider.

Breaking Changes: Simplified Thinking API

Extended thinking (chain-of-thought reasoning) is now a first-class feature in Dartantic with a simplified, unified API across all providers that support thinking:

// OLD
final agent = Agent(
  'openai-responses:gpt5',
  chatModelOptions: OpenAIResponsesChatModelOptions(
    reasoningSummary: OpenAIReasoningSummary.detailed,
  ),
);
final thinking = result.metadata['thinking'] as String?;

// NEW
final agent = Agent('openai-responses:gpt5', enableThinking: true);
final thinking = result.thinking;

• Provider-specific fine-tuning options remain for advanced use cases:
  • GoogleChatModelOptions.thinkingBudgetTokens
  • AnthropicChatOptions.thinkingBudgetTokens
  • OpenAIResponsesChatModelOptions.reasoningSummary

Breaking Change: Removed ProviderCaps

The ProviderCaps type was removed from the provider implementation and moved to a helper function in the tests.

// OLD
final visionProviders = Providers.allWith({ProviderCaps.chatVision});

// NEW
// use Provider.listModels() and choose via ModelInfo instead

New: Custom Headers for Enterprise

All providers now support custom HTTP headers for enterprise scenarios like authentication proxies, request tracing, or compliance logging:

final provider = GoogleProvider(
  apiKey: apiKey,
  headers: {
    'X-Request-ID': requestId,
    'X-Tenant-ID': tenantId,
  },
);

Updated: Google Native JSON Schema Support

Google's Gemini API now uses native JSON Schema support for both:

• Typed output via responseJsonSchema - for structured responses
• Tool parameters via parametersJsonSchema - for function calling

This replaces the previous custom Schema object conversion, enabling better support for complex schemas including anyOf, $ref, and other JSON Schema features that were previously rejected.

This is an internal change with no API surface changes for you except that now you can pass more complex JSON schemas to Google models for both typed output and tool definitions.

New: Google Function Calling Mode

Added functionCallingMode and allowedFunctionNames options to GoogleChatModelOptions for controlling tool/function calling behavior:

final agent = Agent(
  'google',
  chatModelOptions: GoogleChatModelOptions(
    functionCallingMode: GoogleFunctionCallingMode.any, // Force tool calls
    allowedFunctionNames: ['get_weather'], // Limit to specific functions
  ),
);

Available modes:

• auto (default): Model decides when to call functions
• any: Model always calls a function
• none: Model never calls functions
• validated: Like auto but validates calls with constrained decoding

New Model Type: Media Generation

final agent = Agent('google');

// Image generation - uses Nano Banana by default (gemini-2.5-flash-image)
final imageResult = await agent.generateMedia(
  'Create a minimalist robot mascot for a developer conference.',
  mimeTypes: const ['image/png'],
);

// Or specify the model explicitly (like Nano Banana Pro)
final agent = Agent('google?media=gemini-3-pro-image-preview');

• Added media generation APIs to Agent (generateMedia and generateMediaStream) with streaming aggregation helpers.

• Added media generation support for the OpenAIResponsesProvider, GoogleProvider and AnthropicProvider implementations createMediaModel. All three of them support generating media with a prompt and a mime type, using a combination of their intrinsic image generation and their server-side code execution environments to generate files of all types.

• Extended ModelStringParser with media= selectors and added media-specific defaults in the provider registry.

Check out the new media-gen examples to see them in action.

New: Server-Side Tools Across Providers

Server-side tools are now supported across multiple providers:

// Google server-side tools
final agent = Agent(
  'google',
  chatModelOptions: const GoogleChatModelOptions(
    serverSideTools: {GoogleServerSideTool.googleSearch},
  ),
);

// Anthropic server-side tools
final agent = Agent(
  'anthropic',
  chatModelOptions: const AnthropicChatOptions(
    serverSideTools: {AnthropicServerSideTool.webSearch},
  ),
);

You can see how they all work in the new set of server-side tooling examples.

• Anthropic Extended Thinking Support: Added support for Anthropic's extended thinking (chain-of-thought reasoning) exposed in the same way as the OpenAI Responses provider does, so you can write your code to look for thinking output regardless of the provider.
• Ollama Typed Output Support: Ollama now supports JSON schema natively through the updated ollama_dart package.
• Mistral Usage Tracking: The updated mistralai_dart package now includes the usage field natively in ChatCompletionStreamResponse, providing accurate token counts for prompt, response, and totals.
• Cohere Multi-Tool Calling Disabled: Removed ProviderCaps.multiToolCalls from Cohere due to a bug in their OpenAI-compatible API wrt to toolcall IDs.

Another big release!

• Migrated Google provider from deprecated google_generative_ai to generated google_cloud_ai_generativelanguage_v1beta package. This is an internal implementation change with no API surface changes for users. However, it does fix some response formatting issues the deprecated package was having as the underlying API changed; it's so nice to be using the Google-supported package again!
• Internal chat orchestration rearchitecture to simplify Agent and chat model implementations and to focus per-provider orchestration on individual providers.
  • Added Anthropic orchestration provider to handle the toolcall-based method of typed output it requires.
  • Added Google "double agent" orchestrator to support typed output with tools simultaneously. Google's API doesn't support tools and outputSchema in a single call, so the orchestrator transparently executes a two-phase workflow: Phase 1 executes tools, Phase 2 requests structured output. This makes Google functionally equivalent to OpenAI and Anthropic for typed output + tools use cases.
• Restored support for the web! A rogue AI coding agent wrote docs that pulled in dart:io, disabling web support. dartantic_ai fully supports the web and if it ever says it doesn't, that's a bug.
• Used the updated openai_core package to refactor OpenAIResponsesChatModel to eliminate workaround for retrieving container file names.
• Updated the default Anthropic model to claude-sonnet-4-0, although of course you can use whichever model you want.
• Fixed the homepage tag in the pubspec.yaml.
• Added llms.txt and llms-full.txt for LLM readers.

This is a big release!

• Added the OpenAI Responses provider built on openai_core, including session persistence (aka prompt caching), intrinsic server-side tools, and thinking metadata streams. Thanks to @jezell for his most excellent openai_core package and his quick turn-around on my blocking issues!
  • Streaming thinking and server-side tool call progress
  • Server-side image generation with quality, size and partial progress as well as generated image returned as part for ease of access
  • Server-side web search with full progress reports
  • Server-side code interpreter with reusable containers and returning of generated files of all types as parts for ease of access
  • Full server-side vector search tool with example showing how to upload files and query vectors
  • GPT-5 Codex access!
• Replaced the Lambda provider with the new openai_compat.dart sample
• Filtered Cohere models to "Live" entries and updated the default chat model
• Surface usage totals for every provider consistently
• Added support for the DARTANTIC_LOG_LEVEL environment variable for one-line logging configuration
• Support for the fully spec-compliant dotprompt_dart package
• Published all of the specs for dartantic in a new Specifications section in the docs
• 1300+ tests to ensure feature compatibility across the set of supported LLMs

• fix a intermittent anthropic tool-calling error with streaming responses
• fix an openai-based tool-calling error that resulted in an infinite loop from empty responses after tool calls

• move from soti_schema to soti_schema_plus in examples, as the former seems to have been abandoned
• fixed: can't move to the latest version of freezed etc due to deps in ollama_dart #54
• fixed: apiKey and baseUrl parameters should be exposed from the OllamaProvider #52
• fixed: Dartantic AI package local dependency #53
• fixed: update for openai_dart 0.5.4 #51
• fixed chatarang to use new streaming message collection pattern

• Fixed #48: Pass package name and other info to Generative AI providers. I added an example of how to use a custom HTTP client for these kinds of things when creating a model. It's not as easy as it could be, and it didn't work for gemini w/o a quick fix, but it's doable.

• Fixed #47: Dartantic is checking for wrong environment variable. I was being aggressive about constructing providers before they were used and checking API keys before they were needed, which was causing this issue. For example, if you wanted to use Agent('google') and didn't have the MISTRAL_API_KEY set (why would you?), string lookup creates all of the providers, which caused all of them to check for their API key and -- BOOM.

• Updated LLM SDK dependencies:
  • anthropic_sdk_dart: 0.2.1 → 0.2.2
  • openai_dart: 0.5.2 → 0.5.3 (adds nullable choices field support for Groq compatibility)
  • mistralai_dart: 0.0.4 → 0.0.5
  • ollama_dart: 0.2.3 → 0.2.4

• Fixed quickstart example code and updated README

• fixed a compilation error on the web

• updating to dartantic_interface 1.0.1 (that didn't take long : )

Dynamic => Static Provider factories

Provider access has moved to Agent static methods:

// OLD (0.9.x)
final provider = OpenAiProvider();
final providerFactory = Agent.providers['google'];
final providerFactoryByAlias = Agent.providers['gemini'];

// NEW (2.0.0)
final provider1 = Agent.createProvider('openai');
final provider2 = Agent.createProvider('google');
final provider3 = Agent.createProvider('gemini');

If you'd like to extend the list of providers dynamically at runtime, you can use the providerFactories map on the Agent class:

Agent.providerFactories['my-provider'] = MyProvider.new;

Agent.runXxx => Agent.sendXxx

The Agent.runXxx methods have been renamed for consistency with chat models and the new Chat class:

// OLD
final result = await agent.run('Hello');
final typedResult = await agent.runFor<T>('Hello', outputSchema: schema);
await for (final chunk in agent.runStream('Hello')) {...}

// NEW
final result = await agent.send('Hello');
final typedResult = await agent.sendFor<T>('Hello', outputSchema: schema);
await for (final chunk in agent.sendStream('Hello')) {...}

Also, when you're sending a prompt to the agent, instead of passing a list of messages via the messages parameter, you can pass it via the history parameter:

// OLD
final result = await agent.run('Hello', messages: messages);

// NEW
final result = await agent.send('Hello', history: history);

The subtle difference is that the history is a list of previous messages before the prompt + optional attachments, which forms the new message. Love it or don't, but it made sense to me at the time...

Agent.provider => Agent.forProvider

The Agent.provider constructor has been renamed to Agent.forProvider for clarity:

// OLD
final agent = Agent.provider(OpenAiProvider());

// NEW
final agent = Agent.forProvider(Agent.createProvider('anthropic'));

Message => ChatMessage

The Message type has been renamed to ChatMessage for consistency with chat models:

// OLD
var messages = <Message>[];
final response = await agent.run('Hello', messages: messages);
messages = response.messages.toList();

// NEW
var history = <ChatMessage>[];
final response = await agent.send('Hello', history: history);
history.addAll(response.messages);

toSchema => JsonSchema.create

The toSchema method has been dropped in favor of the built-in JsonSchema.create method for simplicity:

// OLD
final schema = <String, dynamic>{
  'type': 'object',
  'properties': {
    'town': {'type': 'string'},
    'country': {'type': 'string'},
  },
  'required': ['town', 'country'],
}.toSchema();

// NEW
final schema = JsonSchema.create({
  'type': 'object',
  'properties': {
    'town': {'type': 'string', 'description': 'Name of the town'},
    'country': {'type': 'string', 'description': 'Name of the country'},
  },
  'required': ['town', 'country'],
});

systemPrompt + Message.system() => ChatMessage.system()

The systemPrompt parameter has been removed from Agent and model constructors. It was confusing to have both a system prompt and a system message, so I've simplified the implementation to use just an optional ChatMessage.system() instead. In practice, you'll want to keep the system message in the history anyway, so think of this as a "pit of success" thing:

// OLD
final agent = Agent(
  'openai',
  systemPrompt: 'You are a helpful assistant.',
);
final result = await agent.send('Hello');

// NEW
final agent = Agent('openai');
final result = await agent.send(
  'Hello',
  history: [
    const ChatMessage.system('You are a helpful assistant.'),
  ],
);

Agent chat and streaming

The agent now streams new messages as they're created along with the output:

final agent = Agent('openai');
final history = <ChatMessage>[];
await for (final chunk in agent.sendStream('Hello', history: history)) {
  // collect text and messages as they're created
  print(chunk.output);
  history.addAll(chunk.messages);
}

If you'd prefer not to collect and track the message history manually, you can use the Chat class to collect messages for you:

final chat = Chat(Agent('openai'));
await for (final chunk in chat.sendStream('Hello')) {
  print(chunk.output);
}

// chat.history is a list of ChatMessage objects

DataPart.file(File) => DataPart.fromFile(XFile)

The DataPart.file constructor has been replaced with DataPart.fromFile to support cross-platform file handling, i.e. the web:

// OLD
import 'dart:io';

final part = await DataPart.file(File('bio.txt'));

// NEW
import 'package:cross_file/cross_file.dart';

final file = XFile.fromData(
  await File('bio.txt').readAsBytes(),
  path: 'bio.txt',
);
final part = await DataPart.fromFile(file);

Model String Format Enhanced

The model string format has been enhanced to support chat, embeddings and other model names using custom relative URI. This was important to be able to specify the model for chat and embeddings separately:

// OLD
Agent('openai');
Agent('openai:gpt-4o');
Agent('openai/gpt-4o');

// NEW - all of the above still work plus:
Agent('openai?chat=gpt-4o&embeddings=text-embedding-3-large');

Agent.embedXxx

The agent gets new Agent.embedXxx methods for creating embeddings for documents and queries:

final agent = Agent('openai');
final embedding = await agent.embedQuery('Hello world');
final results = await agent.embedDocuments(['Text 1', 'Text 2']);
final similarity = EmbeddingsModel.cosineSimilarity(e1, e2);

Also, the cosineSimilarity method has been moved to the EmbeddingsModel.

Automatic Retry

The agent now supports automatic retry for rate limits and failures:

final agent = Agent('openai');
final result = await agent.send('Hello!'); // Automatically retries on 429

Agent<TOutput>(outputSchema) => sendForXxx<TOutput>(outputSchema)

Instead of putting the output schema on the Agent class, it's now on the sendForXxx method:

// OLD
final agent = Agent<Map<String, dynamic>>('openai', outputSchema: ...);
final result = await agent.send('Hello');

// NEW
final agent = Agent('openai');
final result = await agent.sendFor<Map<String, dynamic>>('Hello', outputSchema: ...);

This allows you to be more flexible from message to message.

AgentResponse to ChatResult<MyType>

The AgentResponse type has been renamed to ChatResult.

DotPrompt Support Removed

The dependency on the dotprompt_dart package has been removed from dartantic_ai. However, you can still use the DotPrompt class to parse .prompt files:

import 'package:dotprompt_dart/dotprompt_dart.dart';

final dotPrompt = DotPrompt(...);
final prompt = dotPrompt.render();
final agent = Agent(dotPrompt.frontMatter.model!);
await agent.send(prompt);

Tool Calls with Typed Output

The Agent.sendForXxx method now supports specifying the output type of the tool call:

final provider = Agent.createProvider('openai');
assert(provider.caps.contains(ProviderCaps.typedOutputWithTools));

// tools
final agent = Agent.forProvider(
  provider,
  tools: [currentDateTimeTool, temperatureTool, recipeLookupTool],
);

// typed output
final result = await agent.sendFor<TimeAndTemperature>(
  'What is the time and temperature in Portland, OR?',
  outputSchema: TimeAndTemperature.schema,
  outputFromJson: TimeAndTemperature.fromJson,
);

// magic!
print('time: ${result.output.time}');
print('temperature: ${result.output.temperature}');

Unfortunately, not all providers support this feature. You can check the provider's capabilities to see if it does.

ChatMessage Part Helpers

The ChatMessage class has been enhanced with helpers for extracting specific types of parts from a list:

final message = ChatMessage.system('You are a helpful assistant.');
final text = message.text; // "You are a helpful assistant."
final toolCalls = message.toolCalls; // []
final toolResults = message.toolResults; // []

Usage Tracking

The agent now supports usage tracking:

final result = await agent.send('Hello');
print('Tokens used: ${result.usage.totalTokens}');

Logging

The agent now supports logging:

Agent.loggingOptions = const LoggingOptions(level: LogLevel.ALL);

• Added the ability to set embedding dimensionality
• Removed ToolCallingMode and singleStep mode. Multi-step tool calling is now always enabled.
• Enabled support for web and wasm.
• Breaking Change: Replaced DataPart.file with DataPart.stream for file and image attachments. This improves web and WASM compatibility. Use DataPart.stream(file.openRead(), name: file.path) instead of DataPart.file(File(...)).

• fixed an issue where the OpenAI model only processed the last tool result when multiple tool results existed in a single message, causing unmatched tool call IDs during provider switching.

• Major OpenAI Multi-Step Tool Calling Improvement: Eliminated complex probe mechanism (100+ lines of code) in favor of OpenAI's native parallelToolCalls parameter. This dramatically simplifies the implementation while improving reliability and performance.

• README & docs tweaks

• Completely revamped docs! https://docs.page/csells/dartantic_ai

• Added Agent.environment to allow setting environment variables programmatically. This is especially useful for web applications where traditional environment variables are not available.

• Added support for extending the provider table at runtime, allowing custom providers to be registered dynamically.

• Added optional name parameter to DataPart and LinkPart for better multi-media message creation ergonomics.

• Added ToolCallingMode to control multi-step tool calling behavior.

  • multiStep (default): The agent will continue to send tool results until all of the tool calls have been exercised.
  • singleStep: The agent will perform only one request-response and then stop.

• OpenAI Multi-Step Tool Calling by including probing for additional tool calls when the model responds with text instead of a tool call.

• Gemini multi-step tool calling by handling new tool calls while processing the response from previous tool calling.

• Schema Nullable Properties Fix: Required properties in JSON schemas now correctly set nullable: false in converted Gemini schemas, since required properties cannot be null by definition.

• Breaking Change: McpServer → McpClient: Renamed MCP integration class cuz we're not building a server!
• MCP Required Fields Preservation: Enhanced MCP integration to preserve required fields in tool schemas, allowing LLMs to know what parameters are required for each tool call. This turns out to be a critical piece in whether the LLM is able to call the tool correctly or not.
• Model discovery: Added Provider.listModels() to enumerate available models, and the kinds of operations they support and whether they're in stable or preview/experimental mode.
• Breaking Change: simplifying provider names (again!)
• no change to actual provider aliases, e.g. "gemini" still maps to "google"
• fixed a nasty fully-qualified model naming bug
• Better docs!

• Better docs!

• Breaking change: Content=>List<Part>, lots more List<> => Iterable<>

• Multimedia Input Support: Added attachments parameter to Agent and Model interfaces for including files, data and links.
• Improved OpenAI compatibility for tool calls
• Added the 'gemini-compat' provider for access to Gemini models via the OpenAI endpoint.
• Breaking change: everywhere I passed List<Message> I now pass Iterable<Message>

• Provider Capabilities System: Add support for providers to declare their capabilities
• baseUrl support to enable OpenAI-compatibility
• Added new "openrouter" provider
  • it's an OpenAI API implementation, but doesn't support embeddings
  • which drove support for provider capabilities...
• temperature support
• Breaking change: McpServer.remote now takes a Uri instead of a String for the URL
• Breaking change: Renamed model interface properties for clarity:
  • Model.modelName → Model.generativeModelName
  • Also added Model.embeddingModelName
• Breaking change: Provider capabilities API naming:
  • Provider.caps returns Set<ProviderCaps> instead of Iterable<ProviderCaps>

• MCP (Model Context Protocol) Server Support
• Message construction convenience methods:
  • Added Content type alias for List<Part> to improve readability
  • Added convenience constructors for Message: Message.system(), Message.user(), Message.model()
  • Added Content.text() extension method for easy text content creation
  • Added convenience constructors for ToolPart: ToolPart.call() and ToolPart.result()
• Breaking change: inputType/outputType to inputSchema/outputSchema; I couldn't stand to look at inputType and outputType in the code anymore!
• Add logging support (defaults to off) and a logging example

• Embedding generation: Add methods to generate vector embeddings for text

• Streaming responses via Agent.runStream and related methods.
• Multi-turn chat support
• Provider switching: seamlessly alternate between multiple providers in a single conversation, with full context and tool call/result compatibility.

• added dotprompt_dart package support via Agent.runPrompt(DotPrompt prompt)
• expanded model naming to include "providerName", "providerName:model" or "providerName/model", e.g. "openai" or "googleai/gemini-2.0-flash"
• move types specified by Map<String, dynamic> to a JsonSchema object; added toMap() extension method to JsonSchema and toSchema to Map<String, dynamic> to make going back and forth more convenient.
• move the provider argument to Agent.provider as the most flexible case, but also the less common one. Agent() will contine to take a model string.

• Define tools and their inputs/outputs easily
• Automatically generate LLM-specific tool/output schemas
• Allow for a model descriptor string that just contains a family name so that the provider can choose the default model.

• Multi-Model Support (just Gemini and OpenAI models so far)
• Create agents from model strings (e.g. openai:gpt-4o) or typed providers (e.g. GoogleProvider())
• Automatically check environment for API key if none is provided (not web compatible)
• String output via Agent.run
• Typed output via Agent.runFor

• Initial version.