Shannot

Human-in-the-loop execution for LLM agents.

Shannot lets you see what AI agents want to do before they do it. Scripts run in a supervised environment that captures all commands and file writes. You review in a TUI, approve what's safe, and only then do changes apply.

No more --dangerously-skip-permissions.

How It Works

flowchart LR
    A[🤖 Agent writes script] --> B[shannot run]
    B --> C[📋 Intent captured]
    C --> D[shannot approve]
    D --> E[👤 Human reviews]
    E -->|✓| F[✅ Executes]
    E -->|✗| G[🚫 Blocked]

Loading

Quick Start

# Install
pip install shannot

# Run a script (captures what it wants to do)
shannot run fix-nginx.py

# Review and approve
shannot approve

That's it. Two commands.

Features

Supervised Execution

• All system calls intercepted via PyPy sandbox
• Commands captured during dry-run, executed only after approval
• File writes captured with full content, committed only after approval
• Diff preview for every file modification

Remote Execution

• Run scripts on remote Linux servers via SSH
• Zero dependencies on target — binary deployment
• Same approval workflow, regardless of where code runs

shannot run fix-nginx.py --target admin@prod.example.com
shannot approve

Zero Dependencies

• Pure Python stdlib — nothing to install beyond Python 3.11+
• PyPy sandbox runtime auto-downloads on first use
• Works out of the box on any Linux system

Danger Classification

• Commands color-coded by risk in TUI
• Auto-approve safe operations (ls, cat, df)
• Always-deny destructive patterns (rm -rf /)
• Everything else requires human review

Installation

# Recommended
pip install shannot

# Or with uv
uv tool install shannot

# Or with pipx
pipx install shannot

Requirements:

• Python 3.11+ (host system)
• Linux (sandbox execution) or macOS (remote execution only)

Note: Scripts run in Python 3.6 (the PyPy sandbox version).

CLI Reference

# Core workflow
shannot run <script.py>       # Capture intent
shannot run -c "print(1+1)"   # Inline code
shannot approve               # Review and execute

# Execute specific session
shannot run --session <id>    # Execute approved session
shannot run --session <id> --json-output  # Machine-friendly

# Remote execution
shannot run <script.py> --target user@host

# Setup
shannot setup                 # Interactive menu
shannot setup runtime         # Install PyPy sandbox
shannot setup remote add prod admin@prod.example.com
shannot setup remote test prod
shannot setup mcp install     # Claude Desktop integration

# Status
shannot status                # Runtime, config, pending sessions

Configuration

Single TOML file: ~/.config/shannot/config.toml (or .shannot/config.toml per-project)

[profile]
auto_approve = [
  "ls", "cat", "head", "tail", "df", "ps", "grep", "find",
  "systemctl status", "journalctl",
]
always_deny = [
  "rm -rf /", "rm -rf ~", "dd if=", "mkfs",
  "curl | sh", "wget | bash",
]

[audit]
enabled = true
rotation = "daily"
max_files = 30

[remotes.prod]
host = "prod.example.com"
user = "admin"

[remotes.staging]
host = "staging.local"
user = "deploy"

Why Not Just Use a Container?

Approach	Trade-off
VM/Container	Agent can't do real work — isolated from your actual system
WASM	Capability-restricted — limited to what you expose
Policy sandbox	Static rules — can't adapt to context
Shannot	Agent does real work, with human approval

Shannot is collaborative, not adversarial. The agent helps you. You stay in control.

Use Cases

LLM-assisted sysadmin — Let Claude diagnose and fix server issues, with you approving each change

Safe exploration — Run unfamiliar scripts knowing you'll see exactly what they want to do

Audited automation — Every command and file write logged, nothing happens without approval

Teaching — Show students what scripts do before execution

Security Model

Shannot provides supervised execution, not absolute isolation.

What it provides:

• System call interception via PyPy sandbox
• Virtual filesystem — scripts see only what you expose
• Command and file write approval workflow
• Conflict detection for file modifications
• Audit logging

What it doesn't provide:

• Memory/CPU limits (use cgroups separately)
• Network filtering (sockets are disabled entirely)
• Protection against PyPy sandbox escapes

For production, combine with:

• Dedicated service accounts (least privilege)
• Resource limits (systemd, cgroups)
• Network segmentation

See SECURITY.md for details.

MCP Integration

Shannot includes an MCP server for Claude Desktop:

shannot setup mcp install

This lets Claude propose scripts directly, which you review and approve through the standard workflow.

License

Apache 2.0 — See LICENSE

Contributing

See CONTRIBUTING.md