OpenShell + Pi: build an isolated coding-agent sandbox

A setup guide for running the Pi coding agent inside NVIDIA OpenShell, an isolated environment that limits access to the computer and network.

In plain words
What is it for?
It helps install, configure, migrate, secure, inspect, and troubleshoot Pi with rootless Podman, restricted networking, persistent data, hidden credentials, and agent-to-agent communication.
Why use it?
It reduces the risk of an agent changing the host system, reaching arbitrary websites, or exposing real API credentials.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/howdeploy/deploychan_mcp/openshell-pi-agent
Any agent
npx skills add howdeploy/deploychan_mcp --skill openshell-pi-agent
Clone the repo
git clone --depth 1 https://github.com/howdeploy/deploychan_mcp

Made for: Claude Code, Codex.

Per session 80 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 3,943 The whole file, excluding the scripts and references it only reads on demand.
Security scan C 2 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00080 $0.03943
Opus 5 $0.00040 $0.01972
Sonnet 5 $0.00016 $0.00789
Haiku 4.5 $0.00008 $0.00394

Measured 2d ago against content hash 348c23fd3a46, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade C, and why

OpenShell + Pi: build an isolated coding-agent sandbox scanned grade C with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Downloads and executes remote codehighSupply chain

curl | sh runs whatever the server returns today, which is not necessarily what it returned when this was reviewed.

curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh

Makes network callslowCapability

Not a fault in itself. Listed so you know the mod talks to something, and to what.

curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh
content/skills/openshell-pi-agent/SKILL.md · 521 lines

How it starts

The opening of the file, as written. The whole thing — 521 lines — stays where its author put it; the contents beside it link to each section on GitHub.

OpenShell + Pi

Build a useful Pi coding agent without giving the model ambient access to the host, arbitrary internet egress or real API credentials. Prefer the official Pi community sandbox first; create a custom image only when reproducibility or preinstalled tooling requires it.

This workflow intentionally contains no personal system prompt, private skill, named model vendor or account-specific provider. Select those from the user's own requirements.

Non-negotiable invariants

  • Inspect the current machine and installed CLI before changing anything. OpenShell is fast-moving; use openshell <command> --help as the local command contract.
  • Keep the gateway on loopback with mTLS unless the user explicitly designs a remote gateway topology.
  • Use a rootless compute driver when the host supports it. For Podman, require Podman 5.x, cgroups v2, rootless networking and an active user socket.
  • Keep enable_bind_mounts = false. A host bind mount can negate the filesystem boundary.
  • Keep policy_validation_failure_mode = "fail_closed".
  • Put real credentials in OpenShell provider records, never in --env, images, settings.json, models.json, skills, prompts or project files.
  • Give the sandbox only explicit read-only/read-write paths and explicit egress.
  • Run the agent as a non-root identity and prefer Landlock hard_requirement on a known compatible Linux host.
  • Treat /sandbox as persistent agent state and /tmp as ephemeral. Do not confuse sandbox delete with stop; deletion may remove the managed workspace.
  • Review every Pi extension before loading it. Extensions execute code with the agent user's sandbox permissions.
  • Prove behavior with OS/network evidence. Do not accept “I tested it” from the sandboxed agent without checking the live runtime.

Authorization boundaries

Read-only discovery is safe to perform immediately. Before creating containers, enabling services, replacing policy, attaching credentials, restarting an active agent or changing firewall/network exposure, make sure that action is inside the user's request. Before any delete or workspace replacement, require explicit confirmation and a verified backup.

Read the full file on GitHub · 521 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 521 lines · 80 tokens per session scan C 348c23fd3a46

Subscribe to this mod's changes

OpenShell + Pi: build an isolated coding-agent sandbox is a skill published in the GitHub repository howdeploy/deploychan_mcp (11 stars, last pushed 4d ago), licensed MIT. It adds 80 tokens to every session and 3,943 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it C with 2 findings (downloads and executes remote code, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

portainer-container-compare-chart

Compare two containers using native portainer tool data collection and render SVG/CSV outputs.

rcarmo/piclaw-addons · 24 tokens

docker-sandbox

Create, manage, and execute agent tools (claude, codex) inside Docker sandboxes for isolated code execution. Use when running agent loops, spawning tool subprocesses, or any task requiring process isolation. Triggers on "sandbox", "isolated execution", "docker sandbox", "safe agent execution", or when working on agent…

joelhooks/joelclaw · 74 tokens

browse

Use the browse CLI for Browserbase browser automation, Browserbase cloud APIs, Browserbase Functions, templates, web fetch/search, diagnostics, and Browse.sh skill discovery/installation. Use when the user asks to navigate pages, inspect browser state, run local or remote browser sessions, manage Browserbase…

rivet-dev/agentos · 109 tokens

sanity-check

Run the deferred AgentOS E2E smoke test from public npm packages. Use when the user asks to sanity check, smoke test, or verify a release works.

rivet-dev/agentos · 37 tokens

langbot-deploy

Deploy and configure a LangBot instance — Docker / Docker Compose, Kubernetes, the config.yaml model, the Box sandbox runtime, the plugin runtime, and the global API key. Use when installing, deploying, upgrading, or configuring LangBot in production or self-hosted environments. Triggers on "deploy langbot", "langbot…

langbot-app/LangBot · 104 tokens

launch-graphics

Generate agentOS marketing/launch graphics (launch & social heroes, code-snippet cards) from the committed generators in docs-internal/graphics. Use when asked to make, update, or render an agentOS launch image, social image, "agentOS Apps" graphic, or technical code-snippet image.

rivet-dev/agentos · 66 tokens