VibeThinker — Local AI for Dev Tasks

Run a 3B reasoning model on CPU. Offload quick coding tasks from expensive cloud APIs.

# Install & use in seconds $ pip install vibethinker && vibe setup $ vibe prompt "explain this: git rebase -i HEAD~3" $ vibe code "write a fibonacci generator in rust" -o fib.rs

Local & Private

All inference runs on your machine. Zero data leaves your laptop. No cloud, no tracking, no API keys.

OpenAI Compatible

Drop-in replacement for any OpenAI client. Just point your base_url to http://localhost:8003.

Dev-First

Built-in commands: prompt, chat, code, think, explain, sh, review, eval. Designed for the terminal.

Architecture Overview

VibeThinker is composed of four components: a CLI client, a FastAPI server, an isolated worker subprocess, and a watchdog that keeps everything running.

The server exposes an OpenAI-compatible API on port 8003. The CLI wraps it with developer-friendly commands. In-process model with thread locking and LRU cache (128 entries, 5min TTL) for fast repeated queries.

View full architecture →

100%

core eval pass rate (3/3 tasks)

See eval framework →

Read the API Reference

Installation

# Prerequisites: Python 3.10+, llama-cpp-python
$ pip install vibethinker && vibe setup

# Or from source:
$ git clone https://github.com/bigknoxy/vibethinker
$ cd vibethinker && pip install -e .
   