Jarvis Agent

Your autonomous software engineer, right inside VS Code.

---

Jarvis Agent is a next-generation extension for Visual Studio Code. Whether you want to automate component creation, debug a complex system, or explore a huge codebase, Jarvis integrates seamlessly into your local environment to assist you.

---

Here is an example of what you can do with Jarvis Agent.

✨ Key Features

1. Inline Editing (Cmd+K)

Select any block of code, press Cmd+K Cmd+K (or Ctrl+K Ctrl+K on Windows/Linux), and give Jarvis an instruction (e.g. "Add robust error handling to this function").

The AI writes the code live, and CodeLens (✓ Accept / ✗ Reject) lets you approve or reject the change line by line.

2. Context Understanding (RAG & AST Pruning)

Jarvis doesn't suffer from the typical token limits of other AIs:

• @file: Instantly reference specific files in your requests.
• Semantic RAG (@docs): Search your documentation and entire project by concept, not just by keyword.
• AST Pruning: Jarvis analyzes syntactic structure (with tree-sitter) to read only the important functions of a giant file without saturating the context.

3. Autonomous Execution (Terminal & TDD)

Jarvis doesn't just write code, it runs it.

• TDD Workflow: Type /tdd write me a fibonacci function and Jarvis will write the test, try to run it, see the error, write the implementation, and verify the test passes.
• Integrated Terminal: The agent can run npm install, git commit, or cargo build. Human-in-the-Loop (HITL): in strict mode, it asks for your permission first.

4. Connected to Everything (MCP Standard)

Thanks to the Model Context Protocol (MCP), Jarvis can use external tools. Connect it to GitHub, a PostgreSQL database, or a third-party MCP server directly from the Settings tab. The built-in web search (search_web) is powered by DuckDuckGo — entirely free, no API key or account needed — with configurable default sources (StackOverflow, MDN, GitHub, or your own domains).

5. Specialized Agents & Project Instructions

• 5 dedicated agents (@QA-Agent, @Doc-Agent, @Refactor-Agent, @Security-Agent, @Perf-Agent), each with tool access genuinely restricted to its role — @Security-Agent audits read-only, @QA-Agent runs tests without ever editing code.
• JARVIS.md: type /init (or Jarvis: Initialize Project in the command palette) so Jarvis initializes git if needed, analyzes your project, and generates an instructions file versioned with the code — shared with the whole team, unlike personal rules.
• Folder-scoped rules: scope a rule to a subfolder (src/backend/**) so it only applies to the relevant files.
• Persistent task checklist above the chat, auto-filled for /workflow runs.
• Auto-open of the file being edited, with green/red decorations already in place.
• Inline autocomplete (Tab) — ghost-text suggestions as you type; disabled by default, enable it in Settings > Optimization.

---

🚀 Installation & Quick Start

1. Install Jarvis Agent from the VS Code Marketplace.
2. Click the Jarvis icon in the left sidebar (or run the Jarvis: Start Chat command).
3. The Welcome screen will guide you through configuring your first model!

Provider Compatibility

Jarvis is agnostic. You choose the brain behind your agent:

• Local (Free & Private): Connect Ollama or LM Studio in one click. We recommend recent models like qwen2.5-coder:7b or llama3:8b.
• Cloud (Ultra Performant): Connect OpenRouter, Mistral AI, or any OpenAI-compatible API to use state-of-the-art models. For excellent performance completely free with no latency, we recommend Mistral AI models directly from the Mistral provider. Create a free API key here.

---

📖 Advanced User Guide

Chat Modes

• Automatic: Jarvis analyzes the model's size and the task's complexity to choose between a plain text answer or a decomposition workflow (Task Decomposer).
• Fast: Quick text-only replies, no tool execution.
• Plan: Jarvis always creates a Markdown plan (implementation_plan.md) for you to validate the architecture before it touches any code.

Keyboard Shortcuts

• Cmd+K Cmd+K (Mac) or Ctrl+K Ctrl+K (Win): Inline editing of a code selection.
• Tab: Accepts an inline autocomplete suggestion (if enabled in Settings > Optimization).
• /new: Starts a new conversation (clears context).
• /init: Initializes git if needed and generates JARVIS.md from a project analysis.
• /rollback: Opens the "Checkpoints" list to fully undo Jarvis's impact on your project.

Stop Button

If a task is taking longer than expected or a command is stuck, use the stop button (Square) in the chat panel to interrupt the agent and cancel the current request.

---

⚙️ Advanced Settings & Configuration

All configuration is stored in ~/.jarvis/config.json (+ an optional per-workspace override in jarvis/jarvis-config.json). You can configure Jarvis visually via the Settings tab:

• Adjust verbosity (Concise / Detailed).
• Hide or show the agent's thoughts.
• Manage jarvisignore to prevent access to certain sensitive folders.
• Configure MCP server profiles.
• Configure your default sources for web search (free via DuckDuckGo, no key required).
• Customize specialized agents (system prompt + allowed tools) and rules (global or folder-scoped).
• Enable auto-open of the edited file (always/never/strict-hitl-only) and inline autocomplete (disabled by default).

---

💡 Tips & Best Practices

Get the most out of Jarvis and keep your token usage under control:

• Write a JARVIS.md: run /init once per project. It's versioned with the code (unlike personal rules), so the whole team benefits from the same project context on every request.
• Scope rules to folders: instead of one long global rule, give a rule a scope glob (e.g. src/backend/**) so it only loads for the files it's actually relevant to — shorter prompts, more focused answers.
• Save your own prompts as /shortcuts: Settings > Prompts lets you turn any recurring instruction into a /name shortcut. Jarvis ships with two disabled examples — a Commentator rule and an /explain prompt — enable them or duplicate them as a starting point for your own.
• Reach for @file and @docs before pasting code: @file:path pulls in exactly the file you mean, and @docs:query searches your indexed documentation semantically — both are cheaper and more accurate than dropping large code blocks into the chat.
• Match the chat mode to the task: use Fast for quick text-only questions (no tool calls, no token overhead), Automatic for everyday coding, and Plan when you want to validate the approach before Jarvis touches any files.
• Delegate to a specialized agent: @Security-Agent for a read-only audit, @QA-Agent to just run tests, @Refactor-Agent when you want edits re-verified against the test suite — each has tool access restricted to its role, which keeps it focused and safer to run unattended.
• Let /tdd and /workflow structure multi-step work: /tdd <task> drives a write → test → fix loop automatically, and /workflow <id> <task> runs one of the predefined multi-step sequences (or dynamic for an auto-decomposed task) instead of you prompting step by step.
• Tune HITL to your risk tolerance: strict mode confirms every terminal command and edit, moderate trusts routine actions, free lets Jarvis run autonomously — pick per project depending on how much you want to supervise.
• Use checkpoints as a safety net: Jarvis stashes a git checkpoint before actions/workflows, so /rollback (or the Checkpoints tab) lets you experiment freely and undo anything that goes wrong.
• Keep secrets out of the sandbox: maintain .jarvisignore (auto-generated, stricter than .gitignore) so credentials and sensitive folders are never read or sent to a cloud model.
• Tag model roles deliberately: in Settings > Models, assign a fast/cheap model to autocomplete and a stronger one to chat, instead of using a single model for everything.

---

🤝 Contributing & Support

Jarvis is a community project, built by developers, for developers.

• Ideas, bugs? Open an Issue on GitHub.
• Want to contribute? Pull Requests are welcome! See CONTRIBUTING.md.
• Change history: see CHANGELOG.md.

Made for the VS Code community.