"vi·brisse" (noun): The long, stiff hairs growing around the face of many mammals, used as organs of touch. Whiskers.
Built on the conviction that "Small models + Great tools = Professional performance", Vibrisse transforms local codebases into intelligent conversation partners. It senses patterns and navigates complex architectures with surgical precision.
Vibrisse is not just another LLM wrapper. It's a Sovereign Agent Framework designed for the post-cloud era:
- Privacy as a Feature: Your code never leaves your machine unless you explicitly allow it via Sovereign Routing.
- Efficiency First: We prove that 7B-8B models can outperform GPT-4 on technical tasks when given the right "muscles" (MCP Tools + Surgical RAG).
- Studio-Grade UX: A developer tool should look as good as it works. "Obsidian Glass" isn't just a style; it's a focus-oriented environment.
Vibrisse is a Hybrid Intelligence orchestrator. It intelligently arbitrates between local-first execution (Ollama) and high-capacity cloud models (Groq, OpenRouter) based on task complexity.
- Cost Efficiency: Automatically offloads routine technical tasks (file listing, grep) to local models.
- Privacy Control: You decide exactly when data leaves your machine via a dedicated Sovereign Toggle.
- Heterogeneous Architecture: Decouples the "Router" from the "Worker". Use ultra-fast 270M models (like
functiongemma) for instant intent analysis, reserving heavy 8B/70B models for code generation.
Stop rebuilding what already exists. Vibrisse acts as a centralized MCP Client, allowing you to plug in any tool from the global Model Context Protocol ecosystem:
- Persistent Hub: Connect GitHub, Linear, Slack, or Postgres once; use them across every session.
- Dynamic Tooling: Your agent gains "muscles" in real-time as you add new servers.
For complex tasks, Vibrisse doesn't just output code blindly. It acts as an Architect:
- Human-in-the-Loop: The agent drafts a detailed execution plan and pauses its reasoning graph.
- Interactive UI: The plan is rendered as a clean, interactive
<Artifact>with inline approval buttons. You stay in full control before any destructive action is taken. - Rich Interactive Artifacts: Beyond plans, Vibrisse generates multiple artifact types to enhance UX:
CodeDiff: Visual side-by-side comparison of code changes before writing.Architecture: Mermaid-based interactive diagrams.TaskBoard: Living checklists for long-running tasks.APISpecs: Interactive API testing panels for backend routes.
A premium, immersive Studio experience designed for deep work:
- Thought Graph: Real-time visualization of the agent's reasoning path.
- Thinking Console: Live thought streaming to understand the why behind every action.
- Surgical RAG: Triple-layer retrieval combining Semantic Search, BM25, and high-precision Ripgrep.
- GitHub Alerts: Rich Markdown rendering with warning/info badges (
> [!WARNING]).
Vibrisse is designed to be part of your workflow, not another window you have to manage.
You don't even need to open the UI. Add tags directly in your code comments:
// @vibrisse: refactor this function: The agent will detect the tag in the background and perform the task.# @vibrisse: explain this logic: Get technical explanations injected right where you need them.
In the chat, use triggers to give the agent surgical context:
@Mention Files/Folders: Type@followed by a filename to inject its full content or path into the conversation./Mention Folders: Type/to suggest specific directories for the agent to explore or list.
- Search Everything: Ask "Where is the API key defined?" and Vibrisse will use Ripgrep for 100% accurate results.
- Web Research: Integrated web search via Tavily. No need to touch .env files—simply paste your API key in the Studio Settings.
- MCP Ecosystem: Connect servers to manage Linear tickets, GitHub PRs, or Postgres databases directly from the chat.
- Smart Offloading: If you ask a simple question like "List files in src", Vibrisse will automatically use a Local Model to save you Cloud tokens.
Important
Compatibility: Cross-platform (macOS, Windows/WSL2, Linux). Requires Python 3.12 and Ollama.
Installs Vibrisse as a global command on your system.
macOS / Linux:
# Run the branded one-liner installer
curl -sSL https://agent.vibrisse-studio.dev/install.sh | bashWindows (PowerShell):
# Run the branded one-liner installer
irm https://agent.vibrisse-studio.dev/install.ps1 | iexCommands:
vibrisse: Launch the Web Studio (http://localhost:8001).vibrisse-tui: Launch the terminal Control Center.vibrisse update: Sync latest features and dependencies.
git clone https://github.com/QuentinMerle/vibrisse-agent.git
cd vibrisse-agent
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Dev server (Backend):
uvicorn app.main:app --reload --port 8001
# Dev server (Frontend - Optional):
cd frontend && npm install && npm run devVibrisse is built for and by the local AI community. Whether you are an LLM enthusiast, a React wizard, or a Python architect, your help is welcome!
- Add a Worker: Create specialized agents for specific languages or tasks.
- Connect MCP Servers: Help us build the largest hub of local-first tools.
- Polish the UI: Improve the Obsidian Glass experience.
Check out our Contribution Guide to get started!
Explore the full potential of the Vibrisse ecosystem:
- Capabilities: Explore what the agent can actually do (File I/O, Search, MCP).
- Sovereign Routing: How the hybrid local/cloud arbitration works.
- Ghost Mode: Control the agent via in-file
@vibrisse:directives. - Security & Privacy: Our commitment to local-first data sovereignty.
- Architecture Overview: DDD structure, LangGraph flow, and services.
- Dynamic Personas: Alignment of Wizard-selected personas with specialized system prompts.
- Inference Engines: Guide to Ollama, Custom (vLLM/LM Studio), and Cloud providers.
- Hybrid Retrieval (RAG): How we achieve 100% precision with surgical grep.
- Reasoning & Agents: Supervisor/Worker patterns via langchain-ai/langgraph and robust parsing.
- Setup & Hardware: Requirements for optimal performance.
- Evolution Roadmap: Future vision and upcoming features.
- Agent Strategy: The "Small models, Great tools" philosophy in detail.
- Testing Protocols: How we ensure stability and reliability.
| Command | Action |
|---|---|
/model |
List or change active LLM models |
/path |
Switch target workspace/project |
/scan |
Force architectural re-analysis |
/tools |
List connected MCP servers and tools |
/stats |
Real-time RAM and Ollama health |
Vibrisse AI: Small models, Great tools.
Proudly developed in Beauce, Québec 🇨🇦. Interested in local AI sovereignty? Let's connect!