Skip to main content
Glama
Agentic_SSH

Agentic_SSH

Release Sponsor MCP Toplist agentic_ssh MCP server Crates.io Homebrew

agentic_ssh 🛰️🛡️

Stop your AI agent from nosing around your SSH connections without permission.

A secure-by-default SSH connection pool & MCP server for AI coding assistants (Claude Code, Cursor, Gemini, Antigravity, Copilot, Cline).

agentic_ssh gives your AI assistant fast, token-efficient, and asynchronous SSH access to your homelab, dev servers, and clusters—with strict zero-trust host guardrails so it can never touch unauthorized infrastructure.

AI Agent Multi-Host Interrogation & Security Guardrail

📊 Token & Quota Impact: Raw SSH vs. agentic_ssh

Metric / Scenario

Raw SSH in Agent Prompt

With agentic_ssh MCP

The Difference

Security & Blast Radius

Full access to every host in ~/.ssh/config

Strict Whitelist Guardrail (allow_hosts)

Unauthorized hosts blocked 🛡️

Context Window Consumption

15,000+ tokens (raw stdout dump)

185 tokens (structured JSON telemetry)

98.8% token savings 📉

API Cost per Query

~$0.05 – $0.15

~$0.0004

Fraction of a cent 💰

Connection Latency

800ms – 1.5s reconnect handshake

0ms (pooled Keepalive socket)

Instant response ⚡️

Long-Running Builds

Blocks agent reasoning / silent dropout

Detached async (background: true)

Parallel workflows 🚀


💡 Why I Built This

I built agentic_ssh after watching my AI coding agent do remote server triage across my own fleet of machines.

I asked it to investigate an anomaly. I sat there watching as it opened the same SSH connection time after time to talk to the same remote server, checking one small thing at a time. When querying multiple nodes, it would ask the exact same question of each server individually, running fresh TCP/SSH handshakes for every single command.

Worse, every single docker logs or system check dumped hundreds—sometimes thousands—of tokens of raw terminal noise and ANSI escape codes straight into the context window, chewing through my API budget in minutes. And whenever an idle SSH connection dropped mid-query, the agent would lose its place and spin in retry loops.

I realised that could be dramatically streamlined with a dedicated, connection-pooled, security-first MCP server.


Related MCP server: portal-mcp-server

🛑 Why Giving AI Agents Raw SSH Is a Bad Idea

If you've ever let an autonomous agent run raw terminal ssh, you're exposing your machines to huge risks:

  • 🚨 Unrestricted Fleet Access ("Nosing Around"): Standard agent bash tools have full, unchecked access to your ~/.ssh/config. If you ask an agent to inspect a local dev box, nothing stops it from querying your production database, jumping through client bastions, or running rogue commands where it shouldn't.

  • 💥 The Context Avalanche: The agent runs apt upgrade or cargo build, and 15,000 lines of compiler noise dump straight into your context window—wiping out memory, blowing up rate limits, and wasting money.

  • 👻 Silent Death by Dropout: Cloud NATs and Tailscale love to silently drop idle SSH sockets during a 10-minute compile. The agent hangs forever waiting for output that will never arrive.

  • 😵‍💫 Nested Escaping Hell: Asking an LLM to quote bash inside an SSH string inside an MCP JSON payload (ssh host "bash -c \"echo 'hello'\"") invariably leads to broken quotes and syntax errors.


🪄 How agentic_ssh Fixes This

The Danger

agentic_ssh Zero-Trust Solution

Unchecked Server Access

Strict Host Whitelist (allow_hosts): If allow_hosts is not configured, all hosts are blocked by default. Your agent can only touch machines you have explicitly approved.

Bloated Context Windows

Automatically summarizes verbose outputs and redirects large streams to isolated local session logs (~/.agentic_ssh/sessions/).

Dropped Sockets & Lag

Rust-native connection pooling (russh) with automatic 30s keepalives and zero-latency session reuse.

Broken Quoting

Arguments are structured and escaped natively behind the scenes—zero escaping headaches for the model.

Blocking Long Tasks

Supports background: true—fires long builds or migrations into detached threads so the agent can keep working.


⚡️ Quick Start (30 Seconds)

1. Install

# macOS & Linux (Homebrew)
brew install sandbanks/tap/agentic_ssh

# Or with cargo-binstall (pre-compiled binary)
cargo binstall agentic_ssh

# Or run instantly with Nix (zero compile)
nix run github:sandbanks/agentic_ssh -- doctor

2. Auto-Register with Your AI Agents

One command detects your installed AI tools and registers the MCP server automatically:

agentic_ssh install

(Supports Claude Code / Desktop, Cursor, Gemini, Antigravity, Copilot, Zed, Cline, and Roo-Code).

3. Verify Health

agentic_ssh doctor

🤖 Prompt Recipes: What Your AI Agent Can Do

Once installed, just talk to your agent naturally. Here are real-world prompts you can copy & paste:

🔍 1. Cluster Health & Resource Audit

"Check the CPU load, RAM usage, and available disk space across stan, cartman, and aruba. Report any bottlenecks."

⚡️ Agent calls get_system_stats concurrently across all 3 nodes and gives you a structured comparison table in 2 seconds.

🛡️ 2. Security & Port Exposure Check

"Inspect all active listening TCP and UDP ports on our staging server. Flag anything open on 0.0.0.0 that shouldn't be."

⚡️ Agent calls list_ports with process attribution (PID + binary name) for instant auditing.

🐳 3. Container Status & Error Log Tailing

"Check if any Docker containers crashed on cartman, and tail the last 50 lines of the auth-service logs."

⚡️ Agent calls check_docker_status and tail_container_logs without flooding your context window.

⏳ 4. Detached Background Jobs

"Deploy the latest git commit on stan in the background and notify me when it finishes."

⚡️ Agent runs run_command with background: true, frees up your chat immediately, and tracks the output in a local session log.


🖥️ CLI Superpowers for Humans (Built-in Terminal Tools)

agentic_ssh isn't just an MCP server for AI—it includes powerful CLI tools for human developers too:

💡 Tip: Add alias ash="agentic_ssh" to your ~/.zshrc or ~/.bashrc to save keystrokes!

agentic_ssh watch: Multi-Host Live Streaming TUI

Watch commands run concurrently across multiple servers with live streaming panes and post-run log inspection:

Multi-Host Command Watcher TUI

# Watch a command on multiple hosts concurrently
agentic_ssh watch stan,cartman,aruba "pnpm --version"

# Watch an entire host group defined in your config
agentic_ssh watch web-fleet "docker compose ps"

agentic_ssh json: Instant Multi-Host Scripting

Call any built-in MCP diagnostic tool directly from your terminal and get clean, parseable JSON:

# Get structured system stats across hosts
agentic_ssh json get_system_stats stan,cartman

# Query listening ports
agentic_ssh json list_ports '{"hosts": ["stan", "aruba"]}'

agentic_ssh tui: Live Connection Pool Dashboard

Inspect active SSH sockets, heartbeat metrics, connection lifetimes, and live running commands (with remote PID/PGID and timeout countdowns) inline under each host:

agentic_ssh tui

🧰 Built-in Tool Catalog

MCP Tool

Description

list_hosts

Discovers and returns all authorized SSH host aliases from ~/.ssh/config.

list_groups

Returns defined multi-host server groups (e.g., web-fleet, db-cluster).

run_command

Executes commands concurrently across hosts (supports sync or detached async background mode).

get_system_stats

Fetches CPU load, RAM utilization, and disk partition stats.

list_ports

Scans active listening TCP/UDP sockets with process & PID attribution.

search_processes

Evaluates and filters running processes with regex matching.

tail_log

Safely tails standard system log files without loading multi-gigabyte files.

tail_container_logs

Fetches real-time log frames from Docker containers.

check_docker_status

Returns Docker daemon health, running containers, and image counts.

check_service_status

Queries systemd / systemctl service states.

list_upgradable

Lists pending OS package updates across remote machines.

git_pull

Safely fetches and updates a remote Git repository.

find_large_files

Scans for disk-hogging files exceeding a size threshold.

grep_syslog

Searches remote syslog / journalctl streams for specific error patterns.

list_cron_jobs

Inspects system and user crontabs.

list_network_connections

Lists active network connections and remote endpoints.


⚙️ Configuration (~/.config/agentic_ssh/config.toml)

You can define host groups and security boundaries in an optional configuration file:

# Security boundaries: Only allow safe hosts
allow_hosts = ["stan", "cartman", "kyle", "*.local"]
ignore_hosts = ["prod-db-primary", "secure-vault"]

# Multi-host groups for easy targeting
[groups]
fleet = ["stan", "cartman", "aruba", "kyle"]
pis = ["stan", "cartman", "kyle"]
web = ["aruba", "stan"]

# Custom parameterized agent tools
[tools.deploy_stack]
description = "Pulls latest compose repo and updates stack containers."
command = ["docker", "compose", "-f", "/opt/app/docker-compose.yml", "up", "-d"]
allow_hosts = ["stan", "cartman"]

Remote Shell Requirements & Timeout Process Cleanup

agentic_ssh tracks every active remote command in an in-memory registry and automatically kills timed-out commands (and their spawned child processes) using their remote Process Group ID (PGID).

  • POSIX Shell Requirement: Remote targets are assumed to provide a POSIX-compliant sh or bash (and optionally setsid for isolated process groups).

  • Graceful Fallback: If a target uses a non-POSIX login shell where the PID sentinel cannot be detected, agentic_ssh logs a warning and falls back to best-effort channel closure without failing the command.


🤝 Acknowledgments

  • Standing on the shoulders of giants:

    • Russh: High-performance pure Rust SSH client.

    • Ratatui: Gorgeous terminal user interfaces.

    • Enzo: MCP installer inspiration from tokensave.


💖 Sponsoring Sandbanks

agentic_ssh is built and maintained as independent, sovereign open-source software.

If agentic_ssh saves your token budget, secures your infrastructure, or makes your AI pair-programming workflow faster:

👉 Sponsor @sandbanks on GitHub Sponsors

  • Sponsoring unlocks the GitHub Sponsor badge, access to the private sandbanks/vip discussions and early RFCs, and priority issue triage.


📄 License

Dual-licensed under MIT and Apache 2.0.


MCP Server Identifier: mcp-name: io.github.sandbanks/agentic_ssh

Available Tools

9 tools
get_system_statsA

Fetch CPU load average, RAM, and disk utilization metrics on a single host ('host') or multiple hosts concurrently ('hosts'). If using 'hosts', returns a JSON map mapping hostnames to their metrics. Prefer 'hosts' to query cluster status in parallel.

Inputs: host, hosts.

list_groupsA

Returns a map of configured remote SSH host groups. The map keys represent group names, and the values are lists of target SSH hosts. Useful to see what multi-host SSH groups are available to query/run/watch.

Inputs: no named parameters.

list_hostsA

Returns the list of configured remote SSH hosts. Useful to see what remote machines are available to target.

Inputs: no named parameters.

list_portsA

Lists active listening TCP/UDP ports, matching processes, and PIDs on a single host ('host') or multiple hosts concurrently ('hosts'). If using 'hosts', returns a JSON map mapping hostnames to their active port list. Optionally filter by 'port'. Prefer 'hosts' to scan service availability across multiple machines simultaneously.

Inputs: host, port, hosts.

run_commandA

Executes a shell command on a single host ('host') or multiple hosts concurrently ('hosts'). If using 'hosts', returns a JSON map mapping hostnames to their stdout, stderr, and exit codes. Features optional 'background' execution (starting the command and returning a local log file path immediately for async tracking), 'quiet' execution (suppressing progress logs), and output abbreviation controls. Prefer 'hosts' to execute commands across cluster nodes simultaneously.

Inputs: command (required), host, hosts, quiet, +5 more.

search_processesA

Searches running processes on a single host ('host') or multiple hosts concurrently ('hosts') matching a regex 'pattern'. If using 'hosts', returns a JSON map mapping hostnames to their matched process list. Optionally returns full user/CPU/mem stats if 'full_info' is true. Prefer 'hosts' to find running services across multiple cluster nodes simultaneously.

Inputs: pattern (required), host, hosts, full_info.

tail_container_logsA

Fetch the last N lines of logs from a remote Docker container on a single host ('host') or multiple hosts concurrently ('hosts'). If using 'hosts', returns a JSON map mapping hostnames to their success status and container log output. Prefer 'hosts' to query container logs across multiple machines simultaneously.

Inputs: container (required), host, hosts, lines, +1 more.

tail_logA

Fetch the last N lines of a remote log file on a single host ('host') or multiple hosts concurrently ('hosts'). If using 'hosts', returns a JSON map mapping hostnames to their log output. Prefer 'hosts' to query logs across multiple machines simultaneously.

Inputs: file_path (required), host, hosts, lines.

wait_for_log_patternA

Blocks and streams a remote log file or Docker container logs on a single host ('host') or multiple hosts concurrently ('hosts') until a regex 'pattern' is matched or a timeout is reached. If using 'hosts', returns a JSON map of hostnames to success/error/timeout statuses containing the matched line. Extremely useful for verifying startup or events across cluster nodes without polling.

Inputs: pattern (required), host, hosts, container, +2 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 9 tools

    Disambiguation5/5

    Each tool targets a clearly distinct operation: host/group discovery, system metrics, ports, process search, arbitrary command execution, log tailing, container log tailing, and pattern-based log waiting. Even the two log-tail tools are separated by source type (file vs container), and wait_for_log_pattern has a distinct blocking behavior.

    Naming Consistency5/5

    All tool names follow a consistent snake_case verb-first pattern: list_, get_, run_, search_, tail_, and wait_for_. The verbs are specific and the nouns clearly indicate the target resource, making the naming both predictable and scannable.

    Tool Count5/5

    Nine tools is a well-scoped set for an SSH operations server. It covers host discovery, monitoring, command execution, process inspection, and log observation without bloating the surface or leaving it feeling thin.

    Completeness4/5

    The toolset covers the core SSH workflow: enumerate hosts, inspect system state, run commands, search processes, and monitor logs. The main gap is file transfer or remote file editing, though run_command can work around many of those needs via shell commands.

    Maintenance

    ActivityActive
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      D
      maintenance
      Enables AI agents to establish and manage persistent SSH connections to remote hosts for executing commands. Supports SSH config files, multi-host management, and automatic reconnection with thread-safe concurrent operations.
      15
      51 PyPI
      11
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      SSH orchestration MCP server for coding agents, enabling persistent bash sessions, hash-protected remote file editing, SFTP transfers, SSH tunnels, and multi-host orchestration with connection reuse across tools.
      14
      8
      Apache 2.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to orchestrate a heterogeneous machine fleet via SSH, with unified command execution, file transfer, and dispatch of coding agents across platforms.
      2
      Apache 2.0
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to remotely execute shell commands and transfer files on SSH hosts, including jump-host tunneling, background tasks, resumable transfers, and optional approval of destructive commands, while SSH credentials remain in the local process. It runs over stdio or a streamable HTTP daemon and supports auditing, allow/deny lists, and multiple authentication methods.
      ISC