VibePod CLI – Run AI Coding Agents in Docker

TL;DR

TL;DR: VibePod is a single pip install vibepod away — one CLI to run Claude, Codex, Devstral, Gemini, and other AI coding agents inside isolated Docker or Podman containers, with zero required configuration and a built-in local analytics dashboard.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: vibepod.dev — verified via curl
• Source repository: github.com/VibePod/vibepod-cli — README read end-to-end
• PyPI package: pypi.org/project/vibepod — version 0.19.0, MIT license verified
• HN launch thread: news.ycombinator.com/item?id=47235521
• Docs: vibepod.dev/docs

What Is VibePod?

VibePod is a unified CLI (vp) for running AI coding agents in isolated Docker or Podman containers. It targets developers who want the convenience of a one-command agent runner without managing individual tool configurations or worrying about environment conflicts between agents.

The core promise is zero-config: vp run <agent> just works. The tool handles container provisioning, agent lifecycle, and output streaming while keeping each agent’s environment completely isolated.

Supported agents at the time of writing (from the README):

| Agent | CLI alias | Auto-approval flag | |---|---|---| | Claude | claude | --dangerously-skip-permissions | | Gemini | gemini | --approval-mode=yolo | | Devstral / Vibe | vibe | --auto-approve | | Codex | codex | --dangerously-bypass-approvals-and-sandbox | | Copilot | copilot | --yolo | | Pi | pi | --approve | | Auggie | auggie | — | | OpenCode | opencode | — | | Tau | tau | — | | Jcode | jcode | — | | Antigravity | antigravity | — |

The project is MIT-licensed and written in Python, distributed via PyPI.

Setup Workflow

Prerequisites

• Python 3.10 or later
• Docker or Podman (VibePod auto-detects which is available)

Step 1: Install

pip install vibepod

The package is a single vp executable. No additional dependencies are required — container runtime detection happens at runtime.

Step 2: Run an Agent

vp run claude

To pass arguments through to the agent itself, use -- as a separator so VibePod does not consume them as its own flags:

vp run claude -- --dangerously-skip-permissions "write a hello world in Rust"

Step 3: Use IKWID Mode for Auto-Approval

Agents that support permission-skipping can be launched with --ikwid, which appends the correct flag automatically:

vp run claude --ikwid
vp run codex --ikwid
vp run gemini --ikwid

This is the recommended mode for unattended or scripted runs.

Step 4: Skills — Reusable Prompt Recipes

Install community or project-specific prompt recipes:

vp skills add <skill-name>

Skills are stored per-user or per-project and can be composed into agent invocations. The docs cover the skill format at vibepod.dev/docs/skills.

Step 5: Project Overlays — Custom Base Images

For project-specific toolchains or environment variables, drop a Dockerfile fragment into .vibepod/overlay/. VibePod auto-builds a cached, content-addressed image layer on top of the agent’s base image without requiring a full Dockerfile. See the overlays docs for the full format.

Step 6: Analytics Dashboard

After running agents, open the local dashboard to view usage metrics, HTTP traffic per agent, and token counts:

vp dashboard

All data is stored locally under ~/.vibepod/. Nothing is transmitted to a remote server — the project emphasises privacy-first local collection.

Deeper Analysis

Container Isolation

Each vp run invocation spins up a fresh container. This is the main differentiator from simply aliasing agent CLIs: agents cannot pollute the host environment or conflict with each other’s dependencies. If one agent run goes wrong, the container is discarded and the next run starts clean.

VibePod supports both Docker and Podman transparently. If both are installed, Docker takes priority; this is configurable via the optional YAML config file.

Agent Comparison

The vp compare subcommand (accessible via the dashboard) lets you run multiple agents against the same task and view side-by-side metrics. This is useful for evaluating which agent performs best on a given codebase or problem type.

Herdr Integration

Running vp run inside a herdr pane automatically reports agent state. herdr is a terminal workspace multiplexer — when combined with VibePod, agent output streams directly into a herdr pane alongside your other terminal sessions.

Practical Evaluation Checklist

• [ ] Install via pip install vibepod — verify PyPI package resolves
• [ ] Confirm Docker or Podman is running: docker info or podman info
• [ ] Run vp run claude — verify container starts and CLI streams output
• [ ] Test --ikwid with Claude — verify auto-approval flag is appended
• [ ] Run vp dashboard after a session — verify local metrics file is created
• [ ] Check overlay docs if your project needs custom env vars or tools

Security Notes

• Agent auto-approval flags (--dangerously-skip-permissions, --dangerously-bypass-approvals-and-sandbox, etc.) are exactly as dangerous as they sound. Only use --ikwid in isolated environments you control.
• VibePod does not send telemetry. All metrics stay in ~/.vibepod/ on the host.
• Container networking is not restricted by default — agents can make outbound HTTP requests freely. Restrict network access with your container runtime’s flags if needed.

FAQ

Q: Does VibePod work with local models or only hosted APIs? A: VibePod is designed for CLI-based coding agents that accept command-line invocations (claude, codex, etc.). Whether those agents call local or remote models depends on the agent’s own configuration. VibePod itself does not directly handle model inference.

Q: Can I use VibePod without Docker or Podman installed? A: No. Container isolation is the core feature. If neither runtime is available, vp run will fail with a clear error.

Q: How is this different from just installing agents directly via pip? A: Direct installation puts all agents’ dependencies in the same Python environment, risking conflicts. VibePod isolates each agent run into its own container with a consistent base image, preventing dependency clashes and keeping the host clean.

Q: Is there a way to persist agent state between runs? A: Project overlays (.vibepod/overlay/) let you pre-populate a container with files and environment variables. For full state persistence across runs, mount a host directory as a volume — the docs cover volume mounting in the advanced configuration section.

Q: What is the “IKWID” mode? A: IKWID stands for “I Know What I’m Doing”. It is a convenience flag (--ikwid) that appends the correct auto-approval flag for whichever agent you are running, since each agent uses a different flag name for the same concept.

Conclusion

VibePod is a well-scoped tool that solves a real problem: managing multiple AI coding agents without environment chaos. The zero-config model lowers the barrier to trying new agents, and the built-in analytics dashboard adds immediate value without requiring a separate observability setup.

The MIT license and pure-Python implementation make it easy to inspect, fork, or extend. If you regularly evaluate different coding agents — or want a clean way to run Claude Code alongside Codex in the same project — VibePod is worth a look.

pip install vibepod
vp run claude