Vibrator

vibrator

Run AI coding agents — **Claude Code**, Codex, OpenCode, Kimi Code, or Pi — in an isolated Docker container per project. One Go binary (`vibrate`, alias `vb`), a declarative `.vb` file that pins your setup, and a wizard that fills the gaps.

📖 **[Full documentation →](https://wlame.github.io/vibrator/)**

Claude Code is the primary, fully-exercised harness. Codex, OpenCode, Kimi Code, and Pi are wired in and usable, but **experimental** — expect rough edges and please file issues.

Features

• One binary, zero host clutter — the only thing you install on the host is Docker; everything else runs in the container.
• Isolated per project — a throwaway container per workspace, so your host stays clean and projects never share state.
• Declarative .vb — harness, profile, extensions, mounts, and LLM provider in one TOML file. Gitignored automatically once it holds credentials; an existing .gitignore is always updated. Replayed on every run.
• Wizard fills the gaps — anything you don't pass as a flag, it asks; fully scriptable with --no-wizard.
• Runtime auto-detection — Docker Desktop, OrbStack, Colima, Rancher Desktop, Podman, or native Docker.
• Profiles & features — minimal / backend / frontend / full, fine-tuned with --with / --no.
• Curated extensions — MCP servers, skills, and agents (Serena, Context7, Playwright, claude-mem, the ECC bundle) added with --extensions.
• Language servers on PATH — gopls, basedpyright, and typescript-language-server, baked into the profiles that need them.
• Cloud or local LLMs — Anthropic, OpenAI, any OpenAI-compatible endpoint, or a local Ollama / LM Studio model.
• The container is the boundary — the agent runs with permission prompts bypassed and passwordless root inside, and your workspace is mounted read-write at its real path. That is the design; --no-yolo restores prompts, and --dind removes the boundary entirely.

Install

Download a