mcp_playground_flutter 0.1.5

Interactive AI Agent Playground widget for Flutter. Connect to any LLM provider, register Dart-native tools, remote HTTP MCP servers, and local Node.js or Python subprocesses.

0.1.5

• Upgraded mcp_playground_dart core dependency to 0.1.3.
• Added streaming support refactoring for Small Language Models (SLMs) and on-device models.
• Linked EmbeddedLlmAdapter dynamically using static delegate handlers on LLMService to resolve FFI bindings on desktop platforms.
• Fixed README.md image links to use the correct GitHub raw URLs, resolving broken media rendering issues on pub.dev.

0.1.4

• Upgraded README.md video to embedded test GIF

0.1.3

• Upgraded README.md video to a GIF

0.1.2

• Upgraded mcp_playground_dart core dependency to 0.1.2.
• Added automatic download card rendering for tool output returning base64 binary files (e.g., Excel, PDF) embedded in nested JSON objects.
• Added tool invocation arguments inside the Inspector log messages stream.
• Performance and protocol compliance improvements.

0.1.1

• Bumped version to 0.1.1 for monorepo separation release alignment.
• Fixed WASM compatibility by replacing direct dart:io imports with universal_io.

0.1.0

• Restructured the monorepo, splitting package contents into a pure-Dart core package (mcp_playground_dart) and a Flutter widget library package (mcp_playground_flutter) residing side-by-side to enable clean publication to pub.dev.
• Added export/import functionality for playground saved configurations as JSON files. Export saves all setups to a .json file via FilePicker; import reads a JSON file back, filtering unavailable tools and warning the user via a dialog. Duplicate setups (same name) are skipped on import with a notification.
• Added new localizations for "Export JSON" / "Import JSON" in English and German.
• Enabled web/WASM support for llamadart by removing all kIsWeb blocking guards from EmbeddedLlmAdapter. LlamaBackend() now auto-selects the platform-appropriate backend.
• Added Cross-Origin-Opener-Policy and Cross-Origin-Embedder-Policy meta tags to example web index.html for SharedArrayBuffer support.
• Removed the top-level "Stop after tool call" checkbox from the playground setup screen (now handled per sub-prompt step).

0.0.12

• Added explicit platform declaration in pubspec.yaml to ensure Android, iOS, Windows, macOS, Linux, and Web platform support tags are properly recognized on pub.dev.

0.0.11

• Fixed compile errors in llm_service.dart related to anthropic.InputContentBlock.text missing positional arguments, resolving the 0/20 platform support issue on pub.dev.

0.0.10

• Added user-cancellable prompt execution with a stop button in the chat input bar. When generation is active, the send button is replaced by a red stop-circle button; cancelling skips all remaining sub-prompts and tool iterations.
• Fixed embedded model adapter unhandled-exception noise by replacing the Completer-based load synchronisation with a plain Future<void>? pattern.
• Added context-size fallback retry logic for embedded models: when llama_init_from_model fails, the adapter retries with progressively smaller context sizes (full → half → 1024).
• Improved error messaging in the embedded model picker to suggest alternative quantizations when context creation fails.
• Fixed a Flutter framework assertion in the Agent Inspector panel where ListTile widgets inside a DecoratedBox with a background color lacked a Material ancestor.
• Fixed CreateChartPngTool to accept string numeric values (e.g. "42") from LLM tool-call arguments, preventing type 'String' is not a subtype of type 'num' crashes with embedded models.

0.0.9

• Added embedded model support (discovery from HuggingFace, local GGUF loading, and on-device execution).
• Performed codebase audit and refactoring to optimize services, adapters, and caching.
• Added an embedded example app showcasing offline on-device model configuration.

0.0.8

• Added lenient HTML document parsing and a collapsed "HTML Document Preview" magnifier view modal to robot messages.
• Fixed on-demand tool discovery in the tools catalog dialog by triggering local process connections on dialog load and gracefully shutting down inactive servers upon close.
• Implemented comprehensive translations and localizations support for HTML previews, dialog messages, and tab names in English and German.
• Upgraded Kotlin compiler version in the example app to 2.2.20 to support modern Gradle compilation.
• Filtered base64 image strings from monospace text preview bubbles and LLM request payloads to optimize context tokens.
• Added auto-detection of JSON-embedded spreadsheets and binary files, stripping base64 payloads to show interactive file download cards instead.

0.0.7

• Replaced the standard chat message TextField input bar in the active chat view with SubPromptListEditor, providing the identical multi-prompt entry field, step navigation, and active tools selectors in active chat view.
• Filtered step-specific tool checklist selectors and dropdowns in SubPromptListEditor to show tools belonging exclusively to active/enabled toolsets.
• Fixed auto-enabling behavior for default/initial local MCP servers installed in background post-frame calls on the very first start, fully guaranteeing zero preselected tools on startup when no configurations are loaded.

0.0.6

• Corrected tool state auto-selection on initial app startup: now, if no configuration setup exists, no tools are selected by default.
• Removed the massive 1200+ line hardcoded JSON registry list from mcp_server_registry_tab.dart so it always fetches registry definitions dynamically from the GitHub repository registry.
• Resolved local MCP subprocess connection timeouts and hangs on Windows by invoking process creation with shell integration (runInShell: Platform.isWindows).
• Increased JSON-RPC call timeout to 60 seconds to allow slow local node packages to initialize successfully.
• Added a trailing expand details icon to log tiles in the AgentInspector panel, and refactored log list entries to present properly formatted previews of system prompts, user questions, tool calls, and execution results.

0.0.5

• Implemented multi-prompt step chaining supporting sequential sub-prompt execution, output piping using ${tool_result} / ${task_result} placeholders, per-step active tools filtering, and per-step "stop after tool call" execution flags.
• Added SubPromptListEditor widget to edit multi-prompts on the initial setup screen, featuring desktop side-by-side and mobile stack layouts.
• Refactored active tools selection in playground chat to display only tools belonging to the selected toolset (group) via a dropdown selector, sorting tools alphabetically.
• Provided mobile fullscreen tool selection dialogs using Dialog.fullscreen.
• Added persistent tool selection and server initialization tracking to save active configuration across restarts.
• Updated chat input bar tools selector icon dynamically (all, some, none active tools status states).
• Added active tool schemas (names, descriptions, input schemas) into the system prompt to guide LLMs and improve tool calling correctness.
• Introduced a togglable widget event logging parameter (enableLogging) printing formatted user messages, system prompts, tool calls, and results.
• Wrapped long tool call names in the chat execution bubble headers by removing truncation behavior.

0.0.4

• Improved tool selection dialogs on both the setup screen and chat view by adding per-toolset and global "All / None" toggling controls.
• Prevented toolset groups from disappearing from the chat selection dialog when all their tools are deselected.
• Fixed an issue where external HTTP MCP servers would disappear from the available toolsets list after executing a turn by retaining configured clients regardless of transient connection status.
• Added base64 image and file parsing in chat bubbles to properly render inline image/file attachments (such as PNG charts or spreadsheets) and supported safe, error-free file downloading on Android using scoped storage bytes saving.
• Replaced the hardcoded GitHub MCP server list with an online-fetched list loaded from the remote registry, caching the registry for 24 hours.
• Commited the missing agent walkthrough document back to version control.

0.0.3

• Restructured folder layout to relocate all internal widgets and client submodules under the private lib/src/ directory.
• Added comprehensive dartdoc comments to all exported public API classes, properties, and methods to secure full documentation score points on pub.dev.
• Refined package description metadata in pubspec.yaml to note support for local Node.js and Python subprocesses.

0.0.2

• Upgraded openai_dart to ^7.0.0, anthropic_sdk_dart to ^5.0.0, and ollama_dart to ^2.3.0 and rebuilt all client adapters to match the new SDK specifications.
• Hid the Load, Save, and Clear configuration buttons in the active play/chat view to restrict them only to the initialization screen.
• Upgraded file_picker dependency to ^11.0.2 (migrated to static API).
• Upgraded flutter_lints dependency to ^6.0.0 and cleaned linter syntax.
• Fixed broken visual preview image links in example documentation to use relative paths for pub.dev CDN compatibility.
• Split chart generation examples into distinct "Interactive (fl_chart)" and "PNG Image (canvas)" select tool checkboxes.

0.0.1

• Initial release of mcp_playground_flutter.
• Interactive AI Agent Playground widget with setup and conversation views.
• Multi-LLM support: OpenAI, Anthropic Claude, Google Gemini, Ollama, Mistral AI, and OpenAI-compatible endpoints.
• HTTP MCP Server registry with PulseMCP and Smithery catalog browsing.
• Dart-native local tool system with built-in weather, SSH/SFTP, and chart-generation tools.
• Agentic tool-calling loop with duplicate detection, iteration limits, and safety guards.
• Save/load playground configurations (LLM settings, tool selections, server lists, system prompts).
• Agent Inspector for debugging agent state and conversation history.
• Custom persistence via McpPlaygroundStorageDelegate.
• Cross-platform: Android, iOS, Web, macOS, Windows, Linux.