paude 0.20.0a7

Container wrapper for running AI coding agents in isolated, secure containers

Paude

Run AI coding agents in secure containers. They make commits, you pull them back.

Supported Agents

Agent	Flag	Status
Claude Code	--agent claude (default)	Supported
Codex CLI	--agent codex	Supported
Cursor CLI	--agent cursor	Supported
Gas City	--agent gascity	Supported
Gemini CLI	--agent gemini	Supported
OpenClaw	--agent openclaw	Supported
OpenCode	--agent opencode	Supported

Agents are installed automatically inside the container — no local agent installation needed. You just need authentication credentials for your chosen provider.

Paude-managed agent containers include common development utilities, including rg (ripgrep) for fast code search. The Codex CLI image also includes Node.js for documentation tooling, using the custom base image's package manager or an existing Node.js installation.

Agent installation, credential setup, and provider selection are independent. --agents is the exact install set and its first entry launches as the primary. --providers selects credentials to configure, while --agent-provider maps installed agents to those providers:

paude create \
  --agents gascity,claude,codex \
  --providers vertex,chatgpt \
  --agent-provider gascity=vertex,claude=vertex,codex=chatgpt \
  my-project

Unmapped agents use their default provider. If --providers is omitted, its value is derived from the effective mappings; when supplied, it must include every mapped provider but may include extras. --provider remains shorthand for mapping the primary agent. Gas City no longer installs child CLIs implicitly—list every CLI the image should contain.

Why Paude?

• Isolated execution: Your agent runs in a container, not on your host machine
• Safe autonomous mode: Enable --yolo without fear — the agent can't send your code anywhere
• Git-based workflow: The agent commits inside the container, you git pull the changes
• Run anywhere: Locally or over SSH with Podman or Docker

Demo

The demo shows Claude Code, but the workflow is identical with other agents.

Quick Start

Prerequisites

Container runtime: Podman or Docker, locally or on a remote host reached over SSH.

Authentication — set up credentials for your chosen provider:

Google Cloud / Vertex AI (Claude Code, Gemini CLI, OpenClaw, OpenCode)

Install the Google Cloud SDK, then:

gcloud auth application-default login

Set your project (find the ID in Google Cloud Console):

# Claude Code via Vertex
export CLAUDE_CODE_USE_VERTEX=1
export ANTHROPIC_VERTEX_PROJECT_ID=your-project-id
export GOOGLE_CLOUD_PROJECT=your-project-id

# Gemini CLI / OpenClaw via Vertex
export GOOGLE_CLOUD_PROJECT=your-project-id

# OpenCode via Vertex
export GOOGLE_CLOUD_PROJECT=your-project-id
export VERTEX_LOCATION=us-east5

Anthropic API key (Claude Code, OpenClaw, OpenCode)

export ANTHROPIC_API_KEY=your-api-key

This is the default provider for OpenCode. For Claude Code and OpenClaw, also pass --provider anthropic:

paude create --agent openclaw --provider anthropic ...

OpenAI API key (Codex CLI, OpenClaw, OpenCode)

export OPENAI_API_KEY=your-api-key
paude create --agent openclaw --provider openai ...
paude create --agent opencode --provider openai ...

ChatGPT plan login (Codex CLI)

paude create --agent codex --yolo --git my-project
paude connect my-project

chatgpt is the default provider for the Codex agent.

Inside the session, run codex login. Codex prints a URL and code — complete the login in any browser, on any device (no localhost callback or port forwarding is needed). The session's own proxy sidecar captures the resulting OAuth tokens and manages refresh transparently and independently; real tokens never reach the agent container. Each session has its own OAuth lineage tied to its own private state volume, so multiple ChatGPT-plan sessions can run concurrently without one session's token refresh invalidating another's.

Pass --provider openai to use the plain OPENAI_API_KEY provider instead — no ChatGPT plan, no proxy-managed auth. The default Codex network policy allows chatgpt.com and auth.openai.com for the ChatGPT API and OAuth exchange. Paude merges the required HTTP/SSE provider settings into the session's persistent Codex config.toml because the local MITM proxy does not support the Responses WebSocket transport. Existing model, MCP, project trust, and other user settings are preserved across starts and upgrades. Codex Apps are disabled for ChatGPT OAuth sessions so the Apps MCP integration does not attempt to connect from the container; standalone MCP server configuration is unchanged.

Breaking change from previous versions: paude no longer reads ~/.codex/auth.json from the host. Sessions created under the old host-seeded model must be recreated (paude delete + paude create --agent codex) and re-authenticated with codex login inside the container.

Cursor

agent login  # or set CURSOR_API_KEY=your-api-key

macOS note: On Mac hosts, CURSOR_API_KEY is the simplest authentication method. Without it, each paude session requires a separate browser-based OAuth login via agent login inside the container.

Install

uv tool install paude

First run: Paude pulls container images on first use. This takes a few minutes; subsequent runs start immediately.

Persistent sessions and upgrades

Agent configuration, logins, conversation history, skills, and workspace data are stored on each session's private volume. They survive start, stop, and container recreation. Each session gets a writable /pvc/.gitconfig, seeded from the host config when available or created empty otherwise. All container processes use it, so git config --global works consistently for local and SSH-remote Docker/Podman sessions without modifying the host config. After updating Paude, refresh a session with:

paude upgrade SESSION

The command performs a fresh build with the latest stable agent tooling and migrates state from older containers before replacing them. See Session Management for the per-agent persistence paths.

Your First Session

# OpenClaw — browser-based, no local agent install needed
paude create --agent openclaw --allowed-domains "default openclaw" my-project

# Claude Code (default)
cd your-project
paude create --yolo --git my-project

# Codex CLI
paude create --agent codex --yolo --git my-project

# Cursor CLI
paude create --agent cursor --yolo --git my-project

# Gemini CLI
paude create --agent gemini --yolo --git my-project

# OpenCode
paude create --agent opencode --yolo --git my-project

# Connect to a CLI agent's running session
paude connect my-project

# Pull the agent's commits (use your branch name):
git pull paude-my-project main

You'll know it's working when: For CLI agents, paude connect shows the agent interface and git pull brings back commits. For OpenClaw, paude connect prints a URL — open it in your browser.

OpenTelemetry Export

Export agent telemetry (metrics, logs, traces) to any OTLP-compatible collector:

paude create --otel-endpoint http://collector:4318 my-project

The endpoint hostname is automatically added to the proxy allowlist and non-standard ports (like 4318) are opened in the proxy. Supported agents: Claude Code, Gemini CLI, OpenClaw. Set otel-endpoint in ~/.config/paude/defaults.json to apply globally.

Port Forwarding

Some agents run a service inside the container that you want to reach from your host — for example a web dashboard. Forwarding is opt-in: by default no container ports are exposed. Use --forward-port to forward one or more ports:

# Forward container port 8372 to host port 8372
paude create --forward-port 8372 my-project

# Different host port than container port (HOST:CONTAINER)
paude create --forward-port 9000:8372 my-project

# Bind a non-loopback host interface (HOST_IP:HOST:CONTAINER)
paude create --forward-port 0.0.0.0:8372:8372 my-project

# Repeat the flag to forward multiple ports
paude create --forward-port 8372 --forward-port 5173 my-project

--forward-port accepts three forms:

• PORT — same port on host and container (binds 127.0.0.1)
• HOST:CONTAINER — map a different host port (binds 127.0.0.1)
• HOST_IP:HOST:CONTAINER — bind a specific host interface (e.g. 0.0.0.0 to expose on your LAN)

Forwarding starts when you paude start/paude connect and stops when you disconnect. It works on both the Podman and Docker backends.

macOS (podman machine): Podman runs containers in a VM on macOS, but the forwarder's listener runs on your Mac host directly — it never binds inside the VM. Reaching the container uses the same podman exec path paude connect already relies on, so --forward-port behaves the same as on Linux; no extra podman machine port-publishing is needed.

Loopback binding: the forwarder reaches the service over 127.0.0.1 inside the container, so the in-container service must listen on 127.0.0.1 (or 0.0.0.0) rather than a container-external interface. Every connection appears to the service as coming from 127.0.0.1, which satisfies services that only accept loopback traffic.

You can also set forward-ports in a project's paude.json create section or in ~/.config/paude/defaults.json:

{
  "create": {
    "forward-ports": ["8372"]
  }
}

Resolution follows the usual precedence — CLI --forward-port overrides paude.json, which overrides user defaults (the highest layer that sets any ports wins; layers are not merged).

Worked example: Gas City dashboard

The Gas City agent serves a dashboard inside the container on 127.0.0.1:8372. To reach it from your host:

# Create a Gas City session that forwards the dashboard port
paude create --agent gascity --forward-port 8372 my-project
paude connect my-project

While connected, paude prints Port-forward active: http://127.0.0.1:8372. Open http://localhost:8372 in your host browser to view the dashboard.

Remote hosts (SSH): the forwarder runs on your local machine even when the container runs on a remote host, tunnelling each connection through ssh … podman exec. So the same flag works transparently over SSH — no manual tunnel required:

paude create --agent gascity --forward-port 8372 --host user@remote my-project
paude connect my-project
# Dashboard is reachable at http://localhost:8372 on your LOCAL machine

If you prefer to forward without paude (e.g. for a session created without --forward-port), the equivalent manual SSH tunnel is:

ssh -L 8372:127.0.0.1:8372 user@remote

Note that a plain ssh -L reaches the remote host's loopback, so the container service must additionally be published to the remote host; paude's --forward-port avoids that by exec'ing directly into the container.

Passing a Task

paude create --yolo my-project -a '-p "refactor the auth module"'

File copies also cross the SSH connection automatically. The local path is always resolved on the machine where you run paude:

paude cp dgx-spark-codex:/pvc/workspace/output.log ./output.log
paude cp ./input.txt dgx-spark-codex:/pvc/workspace/input.txt

Or just start the session and type your request in the agent interface.

Something Not Working?

• Run paude --help for all options and examples
• Run paude list to check session status
• Use paude create --dry-run to verify configuration
• Use paude start -v for verbose output (shows sync progress)
• Check credentials: gcloud auth application-default print-access-token (Vertex/Gemini) or verify your API key is exported

---

Learn more:

• Session Management — commands, lifecycle, code sync
• Configuration — defaults, network domains, GitHub CLI, custom environments
• Security Model — attack vectors, --yolo safety, residual risks
• Orchestration — fire-and-forget workflow, harvest, PRs
• Remote Hosts & Docker — SSH remotes, Docker backend, GPU passthrough

How It Works

Your Machine                    Container
    |                              |
    |-- git push ----------------▶ |  Agent works here
    |                              |  (network-filtered)
    ◀-- git pull -----------------|
    |                              |

• Git is the sync mechanism — your local files stay untouched until you pull
• --yolo is safe because network filtering blocks the agent from sending data to arbitrary URLs
• The agent can only reach its API (e.g., Vertex AI) and package registries (e.g., PyPI) by default

Install from Source

git clone https://github.com/bbrowning/paude
cd paude
uv venv --python 3.12 --seed
source .venv/bin/activate
pip install -e .

Requirements

• Python 3.11+ (for the Python package)
• Podman or Docker (for local backend)
• Auth credentials for your provider (Google Cloud SDK, API key, etc.)

Development

See CONTRIBUTING.md for development setup, testing, and release instructions.

License

MIT