OllamaChat 🦙💻

OllamaChat, is a fully native, private, and unlimited AI coding and autonomous software engineering assistant within Visual Studio Code. Your code never leaves your computer, and it requires no API key or subscription.

🌐 English

OllamaChat turns local open-source LLMs (such as DeepSeek-Coder, Qwen2.5-Coder, Llama 3, CodeLlama, Mistral, Phind, etc.) into an integrated pair programmer and autonomous software engineer directly inside your VS Code sidebar.

✨ Key Features

1. 100% Local & Private

• Connects directly to your local Ollama server (http://127.0.0.1:11434 or custom endpoint).
• Zero cloud transmission. Your proprietary codebase, tokens, and data stay strictly on your local machine.
• No rate limits, no paywalls, no monthly subscription fees.

2. Dual Intelligence Modes: Chat & Agent

• 💬 Chat Mode: Clean conversational coding assistant. Read-only mode designed for explanations, algorithm discussions, code reviews, and refactoring tips without making any modifications to workspace files.
• ⚡ Agent Mode: Autonomous coding agent equipped with workspace tools. It can inspect project files, search codebases, create/modify files, organize folders, and run terminal commands to solve tasks end-to-end.

3. Full-Spectrum Agent Toolset

In Agent Mode, models can execute structured workspace operations: | Tool / Action | Description | | :--- | :--- | | Read File (<read_file>) | Inspects file contents for debugging and implementation context. | | Search & Replace (<edit_file>) | Performs surgical, atomic diff replacements in existing files. | | Write File (<write_file>) | Generates or rewrites entire source files. | | Directory Listing (<list_dir>) | Explores folders and directory structure. | | File Search (<find_files>) | Locates files using glob patterns (e.g., *.ts, **/*.py). | | Code Grep (<grep>) | Regex and plaintext line-by-line workspace search. | | Folder Management (<create_folder>) | Recursively creates project directories. | | Delete & Move (<delete_path>, <move_path>) | Safely removes or moves/renames workspace files and directories. | | Terminal Execution (<run_command>) | Runs diagnostic shell commands (e.g., npm test, git status) at the workspace root. |

4. Auto-Apply with Revert & Keep

• File edits are applied immediately for maximum developer velocity.
• Every action card includes:
  • Revert (Geri Al): Undoes file changes back to their original state or removes created files.
  • Keep (Kalıcı Yap): Permanently accepts the changes.
  • View Diff (Farkı Gör): Opens VS Code's side-by-side diff editor.

5. Security Prosecutor (Savcı Modeli)

• Employs a dedicated, ultra-fast lightweight model (e.g., 1.5B or user-defined) to audit every command and file operation in real-time.
• Automatically flags or blocks destructive instructions (such as rm -rf, system-level wipes, unauthorized file overrides).
• Includes an interactive Security Badge & Hover Popover explaining the audit reasoning.

6. Context Mentions (@ & #) and Drag & Drop

• Reference workspace files using @filename or #filename autocomplete.
• Drag and drop files directly from the VS Code Explorer into the chat input to inject them into the context window.

7. Black Box (Kara Kutu) Diagnostic Export

• Click the Black Box button beside any assistant message to instantly generate a comprehensive technical Markdown report:
  • Token metrics: Prompt eval speed, generation speed (tokens/s), duration, and load times.
  • Executed tools, statuses, Prosecutor audit decisions, and output payloads.
  • Complete prompt and thinking trace logs.

8. Model Profiler & KV-Cache Management

• Import locally installed Ollama models with a single click.
• Create custom model profiles with dedicated system prompts, temperature (0.0 - 2.0), KV-cache context windows (num_ctx), and token limits (num_predict).
• Real-time token generation metrics displayed per message.

9. Live Web Search & Internet Access (🌐)

• Toggle live internet access directly from the input toolbar with a single click.
• Models can query the web in real-time to answer questions about modern libraries, updated APIs, and documentation.
• Configurable Analysis Depth:
  • Fast: Title and snippet extraction for lightning-fast responses.
  • Standard: Full page text analysis with HTML stripping and boilerplate cleanup.
  • Deep: Detailed per-page examination and sentence-level synthesis.
• Domain Filtering: Granular domain blacklist (to filter out spam or unwanted sites) and whitelist (to restrict searches strictly to trusted developer docs like MDN, GitHub, StackOverflow).
• Interactive Sources Button (🌐 Sources): Click the Sources button at the bottom of any response that used web information to open a clean Markdown document listing all consulted pages, direct URLs, snippets, and synthesized summaries.
• Autonomous Agent Tool: In Agent Mode, the model can execute <web_search query="..."/> to look up runtime errors and API specifications on the fly.

📦 Installation & Setup

1. Prerequisite: Ensure Ollama is installed and running on your machine:

   ollama serve
   

2. Pull your preferred coding models:

   # Powerful coding models
   ollama pull qwen2.5-coder:7b
   ollama pull deepseek-coder-v2
   
   # Lightweight model for the Security Prosecutor (optional but recommended)
   ollama pull qwen2.5-coder:1.5b
   

3. Install OllamaChat from the VS Code Extension Marketplace.
4. Click the OllamaChat icon in the Activity Bar to begin coding!

⚙️ Configuration & Settings

Configure these options inside the ⚙️ Settings modal:

• Language: Auto (match VS Code language), English, Türkçe.
• Ollama API URL: Local server endpoint (Default: http://127.0.0.1:11434).
• Default System Prompt: Global persona and coding guidelines.
• Max Tokens (num_predict): Unlimited or capped at custom thresholds (minimum 4096).
• Agent Permissions: Automatic approvals for file edits, new file creation, and new folder creation.
• Security Prosecutor: Real-time safety auditor toggle and model selection (auto / 1.5B).
• Web Search Settings: Engine selection (DuckDuckGo, SearXNG), analysis depth (fast, standard, deep), max sources count (1-10), blacklist and whitelist domain rules.