BudAI — your local AI crew

Free & 100% local. Open-source models are the brains. No subscriptions, no API costs — and nothing ever leaves your computer.

BudAI turns free open-source models into the brains of your VS Code — no cloud, no subscriptions, no coding required. Use them to vibe-code (an AI that understands your folder, edits your code and runs commands — you approve every step), build your own AI agents and give them tools, chat with your documents (RAG), search the web, chat directly with any model, and run multi-agent teams — all running 100% on your machine. Just describe what you want and BudAI does the rest.

💙 BudAI is free thanks to support from people like you. If it saves you time, consider sponsoring the developer — it keeps the project alive and independent.

What BudAI can do

✍️ Vibe-code with AI — you stay in control

Chat with your codebase in plain language and let BudAI do the work — no coding required.

• 🔎 It understands your folder. Analyzes the context of your whole workspace — reads and searches your files, finds symbols, and sees the real compiler/linter errors — so it works from your actual code, not a guess. Then it adjusts your code and can run commands in the terminal to get things done.
• 🟥🟩 Smart inline diffs. Every edit shows as a red/green diff right in the editor — removed lines in red, added lines in green — just like a real code review, so you always see exactly what changed.
• ✅ Human-in-the-loop — nothing happens behind your back. ✓ Keep (approve) or ↩ Undo (discard) each edit before it's final, and approve or decline every terminal command before it runs. /code waits for you to accept the change before it runs anything.
• 🛡️ Extra protection built in. Truly destructive commands (like wiping folders) are blocked outright — never even offered, and git commit/push always asks for your approval first, even if you've allowed everything else this session.
• ▶️ Runs & verifies. Runs your scripts and tests, reads the output, and fixes errors. Scripts that need input can prompt you to type answers live in the terminal.

Four modes — pick one, then just describe what you want:

• 💬 /ask — Understand your code. Read-only. Ask anything about your workspace and BudAI reads and searches your files (and the web when needed) to answer in plain language. It never edits or runs anything.
• 📋 /plan — Map out a change. BudAI lays out the steps it would take — without touching a single file. Review the plan, then hit Execute to turn it into real edits.
• 🚀 /code — Make the edits (you're in control). The main coder: it edits files as red/green diffs you ✓ Keep or ↩ Undo, and runs terminal commands only after you approve — and only if you asked it to. Guided, step-by-step, nothing behind your back.
• 🔁 /loop — Keep going until it works. Autonomous mode: BudAI works and verifies round after round — running your scripts/tests and fixing errors — until the goal is actually met (or it hits a safety cap). Great for "just make this work" tasks.

🤖 Build your own AI agents — no code needed

• 🛠️ Describe it or fill a quick form and BudAI builds the agent for you.
• 🧰 Give agents real tools: web search, code execution, web scraping — or add your own custom tools so they can do far more than chat.
• 🧠 Pick any local model from a huge open-source catalog (Llama, Qwen, Mistral, DeepSeek…) — some smarter, some faster — to power each agent.
• 💬 Just talk to them. Chat with one agent right in the panel (@budai /agent <name>), or a plain @budai to chat with any model.

👥 Multi-agent teams

• Put your agents on a team and run them all with a single prompt (@budai /team …) — then choose how they collaborate: Group Chat (a manager routes the discussion), Sequential (an assembly line, each builds on the last), Concurrent (all answer at once), Handoff (they pass control by expertise), or Magentic — great for research, review, or building things together.

💬 Chat directly with any model — with web search

• Type @budai and chat straight with any local Ollama model — no agent or setup needed.
• Turn on web search so the model (or your agents) can pull in current information from the web and cite its sources, instead of relying on stale training data.

📚 Chat with your own documents (RAG)

• Point BudAI at a folder or files and ask questions about your own documents — an LLM, an agent, or a whole team answers from your content, searched and indexed privately on your machine.

🔒 100% private and free

• Everything runs locally through Ollama — zero telemetry, zero API costs, no cloud account, and no data ever leaves your computer.

Powered by Ollama and the Microsoft Agent Framework.

Requirements

• Visual Studio Code 1.104 or newer.
• Python 3.10+ on your PATH — download here (on Windows, tick "Add python.exe to PATH" in the installer). Or point BudAI at a specific interpreter with budai.backend.pythonPath.
• Ollama installed and running — download here.
• At least one local model that supports tools (BudAI needs tool-calling — models are marked tools ✓ in the Models view). For example:
  • General / agents: ollama pull qwen3.5:4b or ollama pull llama3-groq-tool-use
  • Best for vibe-coding: ollama pull qwen2.5-coder:7b — a coding-tuned model writes far cleaner code and follows edit/run steps more reliably. Smaller models (≤4B) work but are noticeably rougher at code.

On first launch BudAI creates its own Python environment in VS Code global storage and installs its backend — a progress notification is shown and this can take a few minutes. Later launches are fast.

Getting Started

1. Install Ollama and pull a model (see above).
2. Install BudAI and click the BudAI icon in the Activity Bar.
3. Chat with a model from the Models view, create an agent in the sidebar Agent Editor, or type @budai in the chat panel.

Settings

• budai.ollama.endpoint: Ollama API endpoint (default http://127.0.0.1:11434).
• budai.backend.pythonPath: Python executable used to create BudAI's backend environment.
• budai.backend.customPort: Backend port (a free port is chosen automatically if busy).
• budai.backend.autoStart: Start the bundled backend automatically when BudAI opens.

Security model

The BudAI backend runs only on 127.0.0.1 and refuses every request that does not come from the extension itself: browser pages are blocked by an Origin check, and all requests must carry a per-launch session token generated by the extension. Agent tools (shell, Python, pip) are powerful by design — they run with your user account, locally — so BudAI adds guardrails in Vibe Code: you approve every terminal command before it runs, catastrophic commands (e.g. rm -rf /, Windows del /s, force-pushing main) are refused outright, and git commit/push always asks first — even under "approve all."

Privacy

BudAI is local-first and collects no usage data. Prompts, files, and generated content are processed by your local Ollama server unless you explicitly use a web feature (web search/fetch). API keys entered in settings are stored in VS Code SecretStorage. The Microsoft Agent Framework's optional telemetry is not configured (traces never leave the process), and BudAI disables even its User-Agent version marker. Any "Code collects usage data" notice you see on VS Code welcome pages refers to VS Code itself and is controlled by VS Code's own telemetry.telemetrySetting.

👋 Want to bring AI to your business? Let's talk.

BudAI is built by an independent AI consultant who helps companies and teams actually adopt AI — not just talk about it. If your business wants to put AI to work — building custom agents, automating workflows, training your team, or just figuring out where to start — I can work with your company directly, hands-on.

👉 Get in touch — let's talk about what AI can do for your business.

Feedback, feature requests & bug reports

Got a feature idea, found a bug, or have a question about the extension? Same place:

• 🌐 machinedragon.github.io/budai/#/contact
• 📧 budai.support@gmail.com

Your feedback shapes what gets built next — every request is read.