Vibes is a high-performance, autonomous TUI (Terminal User Interface) harness designed for small, local LLM models (e.g., Qwen, Gemma, Phi, and bespoke 27B harnesses). It empowers you to manage complex missions by breaking them down into actionable milestones and tasks, all within a sleek, interactive terminal environment.
Built with Ink 6 and React 19, Vibes connects to local LLM servers (Ollama, LM Studio) to provide a state-of-the-art agentic experience with hierarchical planning and DAG-based autonomous scheduling.
- 🎯 Hierarchical Mission Planning: Automatically break down high-level missions into milestones and actionable tasks using the local model's context window.
- 🤖 Autonomous Agent Loop: Executes individual tasks with a robust loop, supporting tool calls, streaming responses, and reasoning-first thinking.
- 📅 DAG-Based Scheduler: Manages task dependencies, parallel execution, deadlock detection for circular
depends_onchains, and auto-discovery of new tasks during execution. - 📟 Interactive TUI: A premium terminal experience featuring:
- Dashboard: High-level progress and statistics.
- Mission View: Tree structure of missions, milestones, and tasks.
- Trace / Task / Review / Logs: Execution trace, live agent output, reviewer feedback, and diagnostic logs.
- Settings: Real-time configuration, model switching, and workspace root (
VIBES_LAUNCH_DIR). - Multiline Mission Prompt: Paste-friendly input with Alt+Enter for newline (works over SSH/Termius).
- 🛠️ Integrated Toolset: Native support for file operations (
read,write,edit,glob,grep), shell command execution, and workspace-wide symbol search. - 🔌 MCP & Plugin Support: First-class support for the Model Context Protocol (MCP) with dynamic environment variable expansion in
.vibes/mcp.json. - ♻️ Context Reconstruction (Beta): Automatically rebuilds a lean, grounded context from state files (
MEMORY.md,checkpoint.md, etc.) when token limits are approached. - 🛡️ Safety & Reliability:
- Auto-Git Snapshots: Commits workspace state before missions; auto-
git initfor scratch workspaces. - Undo Mission (
Alt+Z): Instantly reset the workspace if an agent botches a task. - Thrashing Detection: Automatically pauses if the agent gets stuck in a tool-call loop.
- Scheduler Deadlock Detection: Fails fast with a clear error when the plan has circular task dependencies.
- Auto-Git Snapshots: Commits workspace state before missions; auto-
While Vibes was built to leverage the specific reasoning and tool-calling strengths of Gemma4, its architecture is heavily optimized for local, consumer-grade LLMs. We believe you shouldn't need a cluster of H100s to run a capable agent.
- 🏠 Local-First: Optimized for running on Ollama, LM Studio, or vLLM.
- 📉 Small Model Support: Features like Codex RAG, Local Memory, and JSON Self-Healing allow models as small as 9B (like Qwen 2.5) to perform complex multi-file coding tasks.
- 🎭 Heterogeneous Roles: Assign different models to different roles (Planner, Executor, Reviewer) to balance speed and intelligence.
- 🛡️ The "Executor Contract": A strict verification layer that prevents small models from "hallucinating success" by forcing ground-truth checks against the filesystem.
👉 View our Local LLM & Small Model Benchmark Report for recommended stacks and performance data.
- Node.js: v22 or higher (v20 may work)
- Ollama: Access to an Ollama instance. While Vibes is optimized for Gemma4 26B (
VladimirGav/gemma4-26b-16GB-VRAM:latest), it is compatible with any model that supports OpenAI-style tool calling (e.g., Llama 3.1, Mistral, DeepSeek) by simply updating the.envconfiguration.
The easiest way to install Vibes and set up the global vibes command is using our automated installer:
curl -fsSL https://raw.githubusercontent.com/SPhillips1337/Vibes/main/install.sh | bashThis script will:
- Clone the repository to
~/Vibes. - Install dependencies and build the project.
- Add a
vibesfunction to your shell configuration (.bashrc,.zshrc). - Enable Automatic Workspace Detection (defaults to launch directory; override via
VIBES_LAUNCH_DIRin Settings).
irm https://raw.githubusercontent.com/SPhillips1337/Vibes/main/install.ps1 | iexThis script will:
- Clone the repository to
$HOME\Vibes. - Install dependencies and build the project.
- Add a
vibesfunction to your PowerShell$PROFILE. - Enable Automatic Workspace Detection (Vibes will default to the directory you launched it from).
-
Clone the repository:
git clone https://github.com/SPhillips1337/vibes.git cd vibes -
Install dependencies:
npm install
-
Configure environment: Create a
.envfile in the root directory:OLLAMA_BASE_URL=https://your-ollama-endpoint/v1 OLLAMA_MODEL=VladimirGav/gemma4-26b-16GB-VRAM:latest OLLAMA_API_KEY=ollama CONTEXT_WINDOW=32768
Install the published CLI globally:
npm install --global vibes-tui
vibesOr run it without a global install:
npx vibes-tuiLocal JSONL memory works without additional packages. To use the optional remote Mem0 integration, install it alongside Vibes:
npm install --global vibes-tui mem0aiIf you used the quick installer, you can launch Vibes from any directory on your system:
vibesNote: Vibes defaults to your launch directory as the workspace root. Set Current Working Dir (VIBES_LAUNCH_DIR) in Settings to target a different project folder.
-
Development Mode (with hot reload):
npm run dev
-
Production Build:
npm run build npm start
-
Direct Execution:
tsx src/index.tsx
| Key | Action |
|---|---|
Alt+D |
Switch to Dashboard |
Alt+M |
Switch to Mission view |
Alt+T |
Switch to Trace view |
Alt+Shift+T |
Switch to Task view (live agent output) |
Alt+R |
Switch to Review view |
Alt+L |
Switch to Logs view |
Alt+H |
Switch to History view |
Alt+S |
Open Settings |
Alt+N |
Create a New Mission |
Alt+Enter |
Newline in mission prompt (Enter submits) |
Alt+Z |
Undo Mission (git hard-reset) |
Alt+Y |
Toggle YOLO Mode |
Alt+C |
Toggle Codex RAG |
Alt+X |
Dismiss update notification |
Ctrl+Q |
Quit the application |
Tab |
Cycle focus between panels |
See the full Usage Guide and Configuration in the wiki.
src/index.tsx: Entry point and TUI initialization.src/agent/: The brain of Vibes, containing the mission planner, task executor, and autonomous scheduler.src/tui/: React-based terminal components using Ink.src/tools/: A registry of Zod-validated tools for the agent to interact with the system.
This project is licensed under the MIT License.
Built with ❤️ for the open-source local LLM community.
Test Mission: Create a loading skeleton component with shimmer animation, variant shapes, and Suspense integration Test Concurrency Multislot: You can't parallelize tasks that have logical file dependencies. To test parallel slot utilization, you need a mission with truly independent tasks, e.g.: "Create a React project with a Button component, an Input component, a Card component, and a Badge component — all independent, no shared dependencies."