Safer Ralph

Ralph… but safer.

A bash-based orchestration system for running AI coding agents in Docker containers.

Host: Your machine.
Sandbox: Docker (node, claude, gemini, python, SSH key, git config, port 3000, persistent storage).
Identity: Persona + Sandbox + Keys.
Ralph: The heartbeat loop.

./sandbox.sh build          # Forge Golden Image (once)

go → (ralph)

./sandbox.sh go REPO-URL    # Combines create + clone + in
~/mothership/ralph.sh 1     # Unleash Ralph (inside)

Docs

go split into its component parts, for more control:

create → clone → (ralph) → purge

./sandbox.sh create         # Summon Identity (outputs ID)
./sandbox.sh clone ID URL . # Claim Target
./sandbox.sh in ID          # Jack In
~/mothership/ralph.sh 5     # Unleash Ralph (inside)
./sandbox.sh purge ID       # Scuttle

Container lifecycle: sandbox.sh handles build/create/up/in/purge/list/save with Docker
Identity management: Animal-NATO naming (hawk-alpha, shark-bravo), unique Ed25519 SSH keys per container, emoji Git identities
Agent execution loop: ralph.sh runs inside containers as a heartbeat loop supporting Gemini CLI, Claude Code, and Pi agent (Moonshot)
Personas: Markdown files injected as system prompts (“killer” for YOLO velocity, “step-wise” for single-task precision)
Workspace isolation: Dedicated host directories per container, deploy key injection for GitHub
Credential management: API keys hydrated from env vars on init, never leaked to host

Architecture choice: “Monolithic Script” — bash wrapping CLIs with system prompts. Deliberately minimal: zero infrastructure overhead.

What it doesn’t do: No programmatic SDK integration, no agent-to-agent coordination, no feedback loops beyond ralph.sh iterations, no structured output capture, no state management beyond git commits and progress.txt.