llm_tool 0.7.0

Turn Dart functions into LLM tools for OpenAI, Gemini, Claude and MCP. Annotate with @LlmTool to generate the JSON schema, argument validation and typed calls.

0.7.0

Renamed from llm_tool_calling. This package continues llm_tool_calling (now discontinued) under a shorter name that matches the @LlmTool annotation. The generator is now llm_tool_generator and the Firebase adapter llm_tool_firebase_ai. To switch, update pubspec.yaml and replace package:llm_tool_calling/llm_tool_calling.dart with package:llm_tool/llm_tool.dart. If your build.yaml configures the builder, it's now llm_tool_generator:llm_tool.

Preparing the 1.0 API. Nothing else breaks: old names still work and are marked deprecated until 1.0.0.

• @LlmTool() replaces @Tool(). Almost every AI SDK (firebase_ai, openai_dart, anthropic_sdk_dart, mcp_dart) has its own Tool class, which clashed in files that import both. @Tool() is a deprecated alias.
• toOpenAIJson() and toOpenAIResponsesJson() replace toOpenAiJson() and toOpenAiResponsesJson() (Dart style for two-letter acronyms, like OpenAIClient). The old names are deprecated aliases.
• ToolDefinition<T> is typed by the tool's return type: call returns Future<T>, so results need no cast. Hand-written tools get T inferred.

To upgrade from llm_tool_calling: switch the package names (above), then replace @Tool( with @LlmTool( and toOpenAi with toOpenAI; your IDE shows each remaining place as a deprecation hint.

0.6.2

• README: every "Use with your SDK" example now imports package:llm_tool_calling/llm_tool_calling.dart. toOpenAiJson(), invoke() and the other list methods are extension methods, so a file that only imports your tools.dart can't see them. Added a troubleshooting entry for the resulting error. No code changes.

0.6.1

• Added the package homepage (https://amitgp.dev).

0.6.0

• Built-in support for every major SDK, with no extra packages: toOpenAiJson(), toOpenAiResponsesJson(), toAnthropicJson(), toGeminiJson() and toMcpJson() on a tool or a list of tools (e.g. allTools). Each format is tested against the real SDK (openai_dart, anthropic_sdk_dart, mcp_dart).
• tool.invoke(args, confirm: ...) and allTools.invoke(name, arguments) run a model's call and never throw: validation, confirmation for requiresConfirmation tools, the run, and a JSON-safe ToolResult (toText(), toJson(), isError). Arguments can be a map or a JSON string.
• ToolConfirmation, the callback that asks the user.
• README: "Use with your SDK" with OpenAI, Claude, MCP and Firebase examples (compiled and run in the tests).

0.5.0

• @Param(name: ...) sets the name the LLM sees and sends, e.g. @Param('A game id', name: 'game_id') int gameId.

0.4.0

• withoutAdditionalProperties(schema): a copy of a schema for SDKs that send it to Gemini's older parameters field, which rejects additionalProperties.
• Tool names are documented as up to 63 characters (the firebase_ai limit).
• README: "Use with your SDK" section, with the new llm_tool_calling_firebase_ai adapter.

0.3.0

• No code changes. Released together with llm_tool_calling_generator 0.3.0, which generates a list of all tools in each file.
• README: the list of all tools, and an author section.

0.2.0

• validateArguments() checks nested objects (fields with properties), reporting paths like booking.passengers[0].age is required.
• validateArguments() checks every item of an array against items, reporting paths like tags[2] and grid[1][0].
• validateArguments() checks enum lists and reports the allowed values, e.g. unit must be one of "celsius", "fahrenheit", got "kelvin".

0.1.0

• Initial release.
• @Tool and @Param annotations.
• ToolDefinition with name, description, JSON schema, requiresConfirmation and call(), which validates arguments before running the tool.
• validateArguments(): checks required, unknown and wrongly typed arguments and reports every problem at once. Whole numbers sent as 5.0 are accepted for integer.
• ToolArgumentException with a message an LLM can use to fix its call.