Project Vibe
A Multi-Agent AI Assistant That Thinks, Plans, and Controls Your Desktop
Vibe is a local-first, Jarvis-style AI assistant built around a hierarchical multi-agent architecture. It runs as a native desktop application and combines the reasoning power of large language models with actual computer control — it can browse the web, write and execute code, manage files, click UI elements, and chain these capabilities together to complete complex, multi-step goals autonomously.
Not a Chatbot. Not a Wrapper.
Most “AI assistants” are thin wrappers around a single LLM. Vibe is fundamentally different — it's a Hierarchical Multi-Agent System where a central Managerial Agent orchestrates five specialized agents, each an expert in its domain. This enables parallel task execution, autonomous failure recovery, and real-time agent-to-agent collaboration.
LLM-Agnostic
Gemini, OpenAI, Anthropic, xAI, OpenRouter, Ollama — per agent
Real Desktop Control
Vision + mouse/keyboard — not just generating instructions
Runs Locally
Your data stays on your machine — secrets never leave .env
Fully Extensible
Plugin Development Kit for custom agents and capabilities
6 Agents. One Orchestrator.
Every request flows through the Managerial Agent, which plans, delegates, and synthesizes — while specialized agents handle execution.
The Brain
Managerial Agent
Plans, routes, synthesizes. Breaks complex goals into dependency-resolved subtask graphs, dispatches in parallel where possible, recovers from failures autonomously, and asks clarifying questions mid-turn without ending the session.
Information Specialist
Research Agent
Grounded web search, URL analysis, local file reading, and comprehensive report generation. Google grounding + Brave/DuckDuckGo fallback with concurrent page fetching.
Full-Stack Developer
Coder Agent
Code generation, debugging, refactoring with direct filesystem access. Programmatic Tool Calling (PTC) for batch data tasks — writes and executes scripts instead of ping-ponging through LLM calls. 80-95% token savings.
Vibe's Hands
Computer Use Agent
Full desktop control via vision + mouse/keyboard. DPI-aware coordinate scaling, accessibility API typing, two-tier visual verification, and full goal-completion checks. Runs natively on Windows, macOS, Linux — or sandboxed in a VM.
Visual Creator
Image Agent
AI-powered image generation and analysis. Creates images from text descriptions or edits existing images.
Quality Gate
Code Review Agent
Post-coding review for best practices, security patterns, and efficiency. Auto-triggers based on code complexity or runs on every commit — configurable per session.
Built Different
Every feature exists because we needed it to solve a real problem — not because it looks good on a checklist.
Use Any Model
Gemini, OpenAI, Anthropic, xAI/Grok, OpenRouter, or local Ollama models. Assign different providers per agent. Hot-swap mid-session without losing history.
3-Tier Hybrid Memory
Grep-based fast context, temporal event log, and HNSW vector store — all queried in parallel and fusion-ranked. Per-session and cross-session recall.
Programmatic Tool Calling
For data-heavy tasks, the Coder writes and executes a script instead of burning tokens on LLM round-trips. Auto-routes to host or VM sandbox based on risk.
Agent Collaboration
Research→CUA browser fallback ("Vibe Loop"), Code→CodeReview pipeline, inter-agent clarification requests. Agents talk to each other directly when needed.
VM Sandboxing
Run the Computer Use Agent inside VirtualBox via raw VNC — zero external dependencies. Host, VM, or Auto mode per session. Untrusted scripts auto-route to the sandbox.
Plugin Development Kit
Extend Vibe with custom agents and capabilities. Simple activate(ctx) contract, JSON manifest, and in-app config UI with dynamic schema rendering. No restart needed.
Safety-First Architecture
An AI that controls your desktop needs guardrails. Vibe has a two-layer safety system: a content scanner that blocks dangerous operations at the code level, and an intent-aware verifier that checks whether actions actually match what you asked for.
Path Allowlist/Blacklist — permanent directory-level access control
Command Safety Validator — regex-based BLOCKED / CONFIRM / SAFE for shell commands
Safe Mode / VM-Only Mode — read-only CUA or full sandbox isolation
ENV-Only Key Storage — secrets in .env, only references in config
Log Redaction — API key patterns scrubbed before any log hits disk
A Native Desktop App
Vibe ships as an Electron + Python application — a React GUI connected to a Python backend via ZMQ message bus. Real-time event feed, session dashboard, per-agent configuration, inline code editor, and live CUA status — all in one window.
6
Specialized Agents
6+
LLM Providers
3
Memory Tiers
2
Safety Layers
Interested in Vibe?
Alpha is on the horizon. Get in touch to learn more about licensing, integration, or just to talk multi-agent AI.