ShellGuard

Stop copy-pasting terminal output into your AI. Let your LLM SSH in and look around.

ShellGuard is an MCP server that gives LLM agents read-only bash access to remote servers over SSH. Connect your AI to production, staging, or dev servers and let it run diagnostics, inspect logs, query databases, and troubleshoot -- hands-free.

Commands are restricted to a curated set of read-only tools. Destructive operations are blocked with actionable suggestions so the LLM can self-correct and keep investigating:

• wget -r -> "Recursive downloading is not allowed"
• tail -f -> "Follow mode hangs until timeout. Use tail -n 100 for recent lines."
• sed -> "Stream editing can modify files -- read-only access only. Use grep for searching."
• $HOME/file -> "Variable expansion will not expand. Use absolute paths."

Demo

Here's a demo of ShellGuard with OpenCode against a simulated server. It works extremely well against real servers too.

Quick Start

Install

brew install jonchun/tap/shellguard

Or download the latest binary:

curl -fsSL https://raw.githubusercontent.com/jonchun/shellguard/main/install.sh | sh

Or with Go:

go install github.com/jonchun/shellguard/cmd/shellguard@latest

Configure with an MCP Client

ShellGuard starts as a stdio MCP server -- no arguments needed. Add it to your MCP client of choice:

{
  "mcpServers": {
    "shellguard": {
      "command": "shellguard"
    }
  }
}

{
  "mcpServers": {
    "shellguard": {
      "command": "shellguard"
    }
  }
}

claude mcp add shellguard -- shellguard

{
  "mcp": {
    "shellguard": {
      "type": "local",
      "command": ["shellguard"],
      "enabled": true
    }
  }
}

{
  "mcp": {
    "servers": {
      "shellguard": {
        "type": "stdio",
        "command": "shellguard"
      }
    }
  }
}

{
  "servers": {
    "shellguard": {
      "type": "stdio",
      "command": "shellguard"
    }
  }
}

{
  "context_servers": {
    "shellguard": {
      "command": {
        "path": "shellguard",
        "args": []
      }
    }
  }
}

{
  "mcpServers": {
    "shellguard": {
      "command": "shellguard"
    }
  }
}

What It Does

ShellGuard exposes 7 tools to the LLM:

The LLM connects to a server, runs commands, and reads the output -- the same workflow you'd do manually, but without the context-switching.

How It Works

Every command goes through a pipeline before reaching the remote host:

1. Parse -- bash is parsed into an AST. Shell tricks (semicolons, redirections, command substitution, etc.) are rejected at the syntax level.
2. Validate -- commands, flags, and arguments are checked against a curated allowlist of commands (with an explicit denylist). Default-deny.
3. Reconstruct -- arguments are re-quoted to prevent injection.
4. Execute -- the command runs over SSH with per-command timeouts and output truncation.

For full details, see ARCHITECTURE.md.

SSH Configuration

Authentication

ShellGuard tries authentication methods in this order, stopping at the first success:

Passphrase-protected keys are silently skipped during default key discovery. If you specify a passphrase-protected key via identity_file, the connection will fail. Add the key to your agent first: ssh-add ~/.ssh/my_key.

SSH Modes

ShellGuard supports two SSH modes:

If you connect through bastion hosts, use ProxyJump, or rely on Match blocks in your SSH config, enable system mode:

ssh:
  mode: system

export SHELLGUARD_SSH_MODE=system

If mode is set to system but the ssh binary is not found, ShellGuard logs a warning and falls back to native mode.

Native mode limitations: Match directives, ProxyJump, ProxyCommand, and ForwardAgent are not supported. Use system mode if your infrastructure requires these features.

System mode notes:

• Host key verification is handled entirely by OpenSSH. The host_key_checking and known_hosts_file settings only apply in native mode.
• Connections are multiplexed using OpenSSH ControlMaster, so only the first connection per host pays the SSH handshake cost.

Host Key Verification

ShellGuard verifies SSH host keys using ~/.ssh/known_hosts. Three modes are available:

If a host key has changed, remove the old entry from your known_hosts file.

Configuration

Settings can be specified in a YAML config file or via environment variables. Environment variables take precedence.

Config file location: $XDG_CONFIG_HOME/shellguard/config.yaml (default: ~/.config/shellguard/config.yaml)

ssh:
  mode: "native"                   # native | system
  connect_timeout: "10s"           # default 10s
  retries: 2                       # default 2
  retry_backoff: "250ms"           # default 250ms
  host_key_checking: "accept-new"  # accept-new | strict | off (native mode only)
  known_hosts_file: "~/.ssh/known_hosts"  # native mode only

Toolkit Provisioning

Remote servers don't always have the tools you want. On connect, ShellGuard probes for rg, jq, and yq. If any are missing, the LLM can call provision to deploy them:

Binaries are downloaded from GitHub Releases with SHA-256 verification, cached locally, and deployed to ~/.shellguard/bin/ on the remote host.

Library Usage

ShellGuard can be used as a Go library:

package main

import (
    "context"
    "log/slog"
    "os"

    "github.com/jonchun/shellguard"
)

func main() {
    ctx := context.Background()
    logger := slog.New(slog.NewTextHandler(os.Stderr, nil))

    err := shellguard.RunStdio(ctx, shellguard.Config{Logger: logger})
    if err != nil {
        os.Exit(1)
    }
}

See the Custom Configuration and Custom Executor sections below for advanced usage.

Custom Configuration

import (
    "github.com/jonchun/shellguard"
    "github.com/jonchun/shellguard/manifest"
    "github.com/jonchun/shellguard/server"
)

manifests, _ := manifest.LoadEmbedded()
// Add or remove commands as needed

core, err := shellguard.New(shellguard.Config{
    Manifests: manifests,         // Custom registry (nil = embedded defaults)
    Executor:  myCustomExecutor,  // Custom backend (nil = SSH)
    Name:      "my-server",       // MCP server name
    Version:   "1.0.0",          // MCP server version
})

Custom Executor Backend

Implement the server.Executor interface to use non-SSH backends:

type Executor interface {
    Connect(ctx context.Context, params ssh.ConnectionParams) error
    Execute(ctx context.Context, host, command string, timeout time.Duration) (ssh.ExecResult, error)
    ExecuteRaw(ctx context.Context, host, command string, timeout time.Duration) (ssh.ExecResult, error)
    SFTPSession(host string) (ssh.SFTPClient, error)
    Disconnect(host string) error
}

Testing

make test          # Run all tests
make test-race     # Run with race detector
make lint          # Run go vet

Project Structure

shellguard/
  shellguard.go          # Top-level constructor (New, RunStdio)
  cmd/shellguard/        # CLI entrypoint
  server/                # MCP server core, tool registration, Executor interface
  parser/                # Shell AST parser (mvdan.cc/sh/v3)
  validator/             # Command/flag/SQL validation engine
  manifest/              # YAML command registry (embed.FS)
    manifests/           # allowed command manifests
    manifests/denied/    # denied command manifests
  ssh/                   # SSH manager, ShellQuote, ReconstructCommand
  output/                # Output truncation (64KB cap)
  toolkit/               # Diagnostic tool provisioning (rg, jq, yq)

License

Apache License 2.0. See LICENSE for details.