Privacy-First AI Tool Calling

Summary

  This release introduces AI tool calling with human-in-the-loop approval—a significant step forward for knowledge management workflows. Your AI can now search your vault, read files, and query
   the web, but only with explicit approval at every step.

  Key additions:
  - Vault Search: AI discovers relevant notes across filenames and content
  - File Read: AI requests access to specific files you select
  - Web Search: Brave Search integration (1,000 free queries/month)
  - Editable queries: Modify AI-generated search terms before execution
  - Three-layer approval: Approve execution → Review results → Select what to share

  Architecture Changes

  New service layer (~2,500 lines):
  - ToolService - Orchestrates tool execution and approval workflow
  - VaultSearchService - Full-text vault search via Obsidian API
  - WebSearchService - Brave Search/custom endpoint integration
  - ToolSupportDetector + WhitelistValidator - Model capability detection

  Major refactoring (~22,000 insertions, ~4,500 deletions):
  - Consolidated 6 AI services into unified AiProviderService with provider adapters
  - Replaced ServiceLocator with new ServiceContainer
  - Extracted command handlers from monolithic CommandRegistry
  - Added comprehensive utility modules for error handling, input validation, and message processing

  New UI components:
  - ToolApprovalModal - Shows tool name, parameters, editable queries
  - SearchResultsApprovalModal - Multi-select file filtering
  - WebSearchApprovalModal - Web result filtering with previews

  Privacy & Security

  - Opt-in only: All tools disabled by default
  - Local execution: Vault operations run entirely within Obsidian's API
  - Selective sharing: Choose exactly which results reach the AI
  - No telemetry: Zero external tracking

  Configuration

  Settings → ChatGPT MD → Tool Calling:
  - Enable/disable tool calling (default: off)
  - Brave Search API key (optional)
  - Custom search provider URL
  - Max web results (1-10)

  Model Support

  Tool calling requires model support. The plugin includes:
  - Whitelist of tested models (~150 models across OpenAI, Anthropic, OpenRouter, Gemini)
  - Visual indicator in model selector showing tool compatibility
  - Automatic filtering for unsupported models

  Breaking Changes

  None. Existing configurations work unchanged. Tools are disabled by default.

  Test Plan

  - Verify tool calling toggle enables/disables all tool features
  - Test vault search with various query types
  - Confirm file read approval shows correct file list
  - Test web search with and without API key configured
  - Verify query editing works for vault and web search
  - Test with multiple AI providers (OpenAI, Anthropic, Ollama, etc.)
  - Confirm model selector shows tool support indicators

Author: DenizOkcu

CHANGELOG.md

@@ -1,6 +1,6 @@
 # ChatGPT MD Changelog
 
-## v3.0.0 (December 2025) - Privacy-First AI Tool Calling
+## v3.0.0 - Privacy-First AI Tool Calling
 
 ### 🎯 Major Features
 
@@ -22,6 +22,11 @@
   - Review and filter web results before sharing
   - Custom search provider support for self-hosted solutions
   - Optional full page content fetching
+  - Tool only available to LLM when Brave Search API key is configured
+- **Editable Search Queries**: Users can now edit vault search and web search queries before tool execution
+  - Textarea in tool approval modal for query modification
+  - Real-time validation prevents empty queries
+  - Supports multi-line queries and special characters
 
 ### 🔒 Privacy & Control
 

CLAUDE.md

@@ -1,67 +1,70 @@
 # CLAUDE.md
 
-This file provides guidance to Claude Code when working with code in this repository.
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
 
 ## Project Overview
 
 ChatGPT MD is an Obsidian plugin that integrates multiple AI providers (OpenAI, OpenRouter, Anthropic, Gemini, Ollama, LM Studio) into Obsidian for seamless chat interactions within markdown notes. Users can have AI conversations directly in their notes, with support for note linking, streaming responses, and per-note configuration via frontmatter.
 
-## v3.0.0 Release - Privacy-First AI Tool Calling
+## v3.0.0 - Privacy-First AI Tool Calling
 
-**Latest Release**: v3.0.0 (December 2025)
+Major feature: **Privacy-first AI tool calling** with human-in-the-loop approval:
 
-Major new feature: **Privacy-first AI tool calling** with human-in-the-loop approval:
 - **Vault Search**: AI can search your notes (you approve which files to share)
 - **File Reading**: AI can request access to specific files (you select which ones)
 - **Web Search**: AI can search the web via Brave Search API (1,000 free queries/month)
 - **Three-Layer Approval**: Approve execution → Review results → Select what to share
-- **All tools disabled by default** (opt-in feature only)
-
-See [`planning/code-review/`](planning/code-review/) for comprehensive code review and implementation guidance.
+- **All tools disabled by default** (opt-in via Settings → ChatGPT MD → Tool Calling)
 
 ## Quick Reference
 
 **Entry point**: `src/main.ts` → `main.js`
 
-**Common commands**:
+**Commands**:
 
 ```bash
 npm run dev        # Development with watch mode
 npm run build      # Production build with TypeScript checks
 npm run lint       # Check code quality
 npm run lint:fix   # Auto-fix linting issues
+npm run analyze    # Bundle size analysis
 ```
 
+**No test suite**: This project does not currently have automated tests.
+
 ## Architecture Overview
 
-The plugin uses **Service Locator pattern** for dependency injection:
+The plugin uses **constructor injection** via `ServiceContainer`:
 
-- `src/core/ServiceLocator.ts` - Central DI container
-- `src/core/CommandRegistry.ts` - Manages all Obsidian commands
-- AI services implement `IAiApiService` interface
+- `src/core/ServiceContainer.ts` - DI container with readonly service instances
+- `src/Commands/` - Command handlers (extracted from old CommandRegistry)
+- `src/Services/AiProviderService.ts` - Unified AI service with adapter pattern
 
-**Message flow**: User command → EditorService extracts messages → MessageService parses → AI service calls API → Response streamed to editor
+**AI SDK**: Uses Vercel AI SDK (`ai` package) with provider-specific adapters (`@ai-sdk/openai`, `@ai-sdk/anthropic`, `@ai-sdk/google`, `@openrouter/ai-sdk-provider`).
+
+**Message flow**: User command → EditorService extracts messages → MessageService parses → AiProviderService calls API → Response streamed to editor
 
 ## Code Organization
 
-Each directory has its own CLAUDE.md with detailed context that auto-loads when you work in that area:
+Each directory has its own CLAUDE.md with detailed context:
 
-- `src/core/` - Core infrastructure (ServiceLocator, CommandRegistry)
-- `src/Services/` - Service implementations
-- `src/Views/` - UI components
+- `src/core/` - ServiceContainer (DI)
+- `src/Commands/` - Obsidian command handlers
+- `src/Services/` - Service implementations + `Adapters/` subdirectory
+- `src/Views/` - UI components and modals
 - `src/Models/` - TypeScript interfaces
+- `src/Types/` - AI service type definitions
+- `src/Utilities/` - Pure helper functions
 
 ## Cross-cutting Documentation
 
-For topics that span multiple areas:
-
 - **[docs/development.md](docs/development.md)** - Build process, tooling, esbuild setup
 - **[docs/message-flow.md](docs/message-flow.md)** - Complete flow from user input to AI response
 
 ## Key Design Patterns
 
-1. **Service Locator** - Centralized dependency injection
-2. **Strategy Pattern** - Different AI services, same interface
+1. **Constructor Injection** - Dependencies passed via ServiceContainer
+2. **Adapter Pattern** - Provider-specific adapters implement common interface
 3. **Frontmatter-driven config** - Per-note settings override globals
-4. **Streaming responses** - Real-time AI output with SSE
+4. **Streaming responses** - Real-time AI output via Vercel AI SDK
 5. **Link context injection** - Auto-include `[[Wiki Links]]` in prompts

README.md

@@ -1,45 +1,15 @@
 # ChatGPT MD
 
-🚀 A seamless integration of ChatGPT, OpenRouter.ai and local LLMs via Ollama into Obsidian.
+🚀 A seamless integration of ChatGPT, OpenRouter.ai and local LLMs via Ollama/LM Studio into Obsidian.
 
 ![Chatting with links about vacation plans](images/chat-with-link.gif)
 
-## 🚀 What's New in v3.0.0: Privacy-First AI Tool Calling
+## 🚀 What's New in v3.0.0: Privacy-First AI Tool Calling (off by default - Settings → ChatGPT MD → Tool Calling)
 
-**Your AI assistant can now search your vault, read files, and search the web—with explicit approval at every step.**
+**Your AI assistant can now actively search your vault, read files, and query the web—with a human-in-the-loop architecture that keeps you in control.**
 
-### Key Features
+v3.0.0 introduces a **tool calling system** built on privacy-first principles. When your AI needs information, it requests permission to use tools—you approve execution, review results, and control exactly what gets shared back to the model. Nothing leaves your vault without explicit consent.
 
-- **🔍 Vault Search**: AI discovers relevant notes in your vault (you approve which files to share)
-- **📄 File Reading**: AI can request access to specific files (you select which ones)
-- **🌐 Web Search**: AI searches the web via Brave Search API (1,000 free queries/month)
-- **✅ Three-Layer Approval System**: 
-  1. Approve what the AI wants to do
-  2. Review what it found
-  3. Select exactly which results to share
-- **🔒 Privacy-First Design**: No data reaches the AI without your explicit consent
-- **🎯 All Providers Supported**: OpenAI, Anthropic, Gemini, OpenRouter, Ollama, LM Studio
-- **⚙️ Disabled by Default**: Opt-in feature only—enable in Settings → Tool Calling
-
-### Why This Matters
-
-For privacy-conscious note-takers, this is a game-changer. You get the power of AI with full control over your data. Your vault stays yours.
-
-### Getting Started with Tool Calling
-
-1. **Install the beta**: Use [BRAT plugin](https://github.com/TfTHacker/obsidian42-brat) and select version **2.12.0-beta**
-2. **Enable in settings**: Go to Settings → ChatGPT MD → Tool Calling
-3. **Optional**: Add Brave Search API key for web search (1,000 free queries/month)
-4. **Start chatting**: Your AI will ask for approval when it wants to use tools
-
-⚠️ **Beta Warning**: Test on a backup/test vault first. This is a beta release with active development.
-
-### Use Cases
-
-- **Research Assistant**: "Search my vault for notes about quantum computing and find recent papers"
-- **Knowledge Synthesis**: "Find all my Q3 meeting notes and summarize key decisions"
-- **Web-Enhanced Writing**: "Search the web for latest statistics and incorporate them"
-- **Note Discovery**: "Find connections between notes I might have missed"
 
 ## A simple and quick Start 🏁
 Get started in just a few simple steps:
@@ -160,6 +130,77 @@ openaiUrl: https://api.openai.com
 💡 Pro tip: Increasing `max_tokens` to a higher value e.g. `4096` for more complex tasks like reasoning, coding or text creation.
 The default model `gpt-5-mini` is optimized for speed and efficiency. Upgrade to `gpt-5` for enhanced reasoning capabilities or use `gpt-5-nano` for ultra-lightweight responses.
 
+### Tools
+
+1. **Install**: Update or install v3.0.0+ from Obsidian
+2. **Configure**: Settings → ChatGPT MD → Tool Calling → Enable,
+3. **Optional**: Add [Brave Search API key](https://brave.com/search/api/) (free tier: 1,000 queries/month)
+4. **Chat**: Use the `ChatGPT MD: Chat` command. AI will request tool use when needed.
+
+The implementation follows a three-layer approval pattern:
+
+1. **Execution Layer**: AI requests tool use with parameters
+2. **Processing Layer**: Tool executes locally in your vault using Obsidian's API (full-text search across filenames and content)
+3. **Approval Layer**: Interactive modals let you filter results before they're returned to the AI
+
+### Available Tools
+
+**Vault Search** (`vault_search`)
+- Multi-word OR search: matches ANY query term across your vault
+- Searches both filenames and file content simultaneously
+- Excludes current file to prevent recursion
+- Configurable result limits (default: 5 files)
+- Query editing: refine search terms before execution
+
+**File Read** (`file_read`)
+- Direct file access when AI knows specific file paths
+- Batch reading support for multiple files
+- Full content extraction with your approval
+- Useful for targeted lookups once files are discovered
+
+**Web Search** (`web_search`)
+- Powered by Brave Search API (privacy-focused, 1,000 free queries/month)
+- Custom search provider support for self-hosted endpoints
+- Optional full-page content fetching
+- Automatic API key validation—tool only appears when configured
+- Query editing: modify web search queries before execution
+
+### Privacy & Security
+
+- **Local-First Execution**: All vault operations run entirely within Obsidian's API
+- **Selective Sharing**: Multi-select modals let you choose exactly which results to share
+- **No Telemetry**: Zero tracking or analytics—tool usage stays private
+
+### Configuration
+
+Enable tool calling in **Settings → ChatGPT MD → Tool Calling**:
+
+- **Enable Tool Calling**: Master switch (default: disabled)
+- **Brave Search API Key**: Your Brave Search API key
+- **Custom Provider URL**: Self-hosted search endpoint
+- **Max Web Results**: Number of web results to return (1-10)
+
+### Use Cases
+
+**Research Assistant**: "Search my vault for notes about quantum computing algorithms and recent papers on the topic"
+
+→ AI discovers relevant notes → You approve which files to share → AI synthesizes information with proper attribution
+
+**Knowledge Synthesis**: "Find all my Q3 meeting notes and summarize key decisions about product roadmap"
+
+→ Vault search returns meeting files → You select the relevant ones → AI extracts and summarizes decisions
+
+**Web-Enhanced Writing**: "Search the web for latest climate change statistics and incorporate them into my article"
+
+→ Web search fetches current data → You filter reliable sources → AI integrates citations into your draft
+
+**Cross-Reference Discovery**: "Find notes that mention both machine learning and productivity techniques"
+
+→ Multi-word OR search finds intersections → You approve interesting connections → AI highlights patterns you might have missed
+
+⚠️ **Note**: Tool support depends on model capabilities. Older models may not support function calling. You can check tool capabilities in the tool selection list after enabling tool support in the settings.
+
+
 ### Multi Model Chats
 You can set and change the model for each request in your note. 
 Specify the `model` property via frontmatter: