Servonaut MCP Server

Servonaut

Your servers. Your terminal. Your AI agent. One TUI.

Manage AWS, Hetzner, OVH, and custom servers from one terminal — with a built-in AI assistant and MCP server.

Quick Install

Linux / macOS:

curl -sSL https://raw.githubusercontent.com/zb-ss/servonaut/master/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/zb-ss/servonaut/master/install.ps1 | iex

Or install directly via pipx / pip:

pipx install servonaut

Manual install from source:

git clone https://github.com/zb-ss/servonaut.git
cd servonaut
pipx install .

Set up with an AI agent

Prefer to let an AI agent do the whole thing? Paste this prompt into Claude Code, Cursor, or any coding assistant — it installs Servonaut, generates the config, and walks you through AWS / SSH / bastion / custom-server / AI-provider setup plus the MCP server.

Install and configure Servonaut, a TUI for managing servers (AWS EC2, OVHcloud, Hetzner Cloud, and custom SSH servers).

1. Install with all optional features: `pipx install 'servonaut[all]'`
   (bundles the MCP server + OVH/Hetzner SDKs + keyring; AI log analysis needs no extra. Use plain `pipx install servonaut` for a minimal install.)
2. Run `servonaut` once to generate ~/.servonaut/config.json
3. (Optional) If I have a Servonaut account, run `servonaut login` to unlock the hosted features: Servonaut AI (chat with my fleet, no local API key), config sync across machines, Memory Sync, and proactive monitoring (Findings). Servonaut works fully offline against my own credentials if I skip this.
4. Read ~/.servonaut/config.json and help me configure:
   - AWS regions to scan (default scans all, set `regions` array to limit)
   - Default SSH username (`default_username`, default "ec2-user")
   - Cache TTL (`cache_ttl_seconds`, default 3600)
   - Terminal emulator if not auto-detected (`terminal_emulator`)
5. If I use bastion/jump hosts, help me set up `connection_profiles` and `connection_rules`
6. If I have non-AWS servers, help me add them to `custom_servers`
7. If I use OVHcloud or Hetzner Cloud, help me add the API credentials so those instances merge into the fleet
8. For AI log analysis or chat with my own model (instead of Servonaut AI), help me configure `ai_provider` (openai/anthropic/gemini/ollama)
   - Each provider has its own key field (`openai_api_key`, `anthropic_api_key`, `gemini_api_key`, `ollama_api_key`); local Ollama needs none
   - Key fields support `$ENV_VAR` and `file:~/.secrets/key` syntax so secrets stay out of the config file
9. Install the MCP server into my coding agent: `servonaut --mcp-install claude` (or `cursor`, `windsurf`, `opencode`, `vscode`, `all`)
10. (Optional) To let AI agents/teammates reach this machine over the relay — and to run proactive Findings scans — start it with `servonaut connect`

After setup, launch with `servonaut` and walk me through the key features, including the Findings inbox if I enabled the hosted features.

Screenshots

Instance list — AWS, Hetzner, OVH, and custom servers merged into one view

Sidebar reveals Fleet Memory, Memory Sync, Secrets, Settings, and per-provider management for OVH and Hetzner

Built-in AI assistant with MCP server integration — chat with local providers or hosted Servonaut AI

CloudWatch log browsing with Top IPs analysis, geolocation, and abuse scoring

Ban/unban IPs via WAF, Security Groups, or NACLs with audit trail

All screenshots and the launch video were recorded with --demo active, which replaces real IPs, ARNs, paths, and secrets with safe fake equivalents. See docs/demo-mode.md for what is redacted and how to use it.

Features

Badges: Solo+ = included with paid Solo/Teams plans.

Core & connectivity

• Interactive TUI — mouse + keyboard, powered by Textual.
• Multi-provider fleet — AWS EC2, OVHcloud (dedicated / VPS / Public Cloud), Hetzner Cloud, and custom servers from any provider (DigitalOcean, on-prem, …) — listed and searchable in one view across all regions.
• Per-instance dashboard — click a server for a Server Actions view: a memory snapshot (OS, disk, web stack, databases, runtimes, containers) plus an opt-in live resource monitor (L — CPU / RAM / load / disk / uptime, polled only while open).
• SSH & SCP — one-key SSH in a new terminal window (auto-detected emulator); upload/download files and directories.
• Run remote commands — overlay panel with real-time streaming output, history, and saved favorites.
• Remote file browser — interactive file-tree navigation, inline in the dashboard or full-screen.
• Real-time log viewer — stream logs via tail -f with pause, search, and log switching.
• Robust SSH — bastion / jump-server (ProxyJump / ProxyCommand), keepalives on by default (tunable), per-host extra_ssh_options for legacy boxes, and key auto-discovery.

Cloud provider management

• OVHcloud — OVH → ⚙ Manage: create / start / stop / reboot / delete (Cloud / VPS / dedicated), a region-first create wizard with API-backed pricing, plus DNS, IP blocks & failover IPs, snapshots, block storage, and billing.
• Hetzner Cloud — Hetzner → ⚙ Manage: full lifecycle + project SSH-key registry, with an equivalent CLI (servonaut hetzner …). Auto-registers new servers. → docs

Observability & security

• Proactive monitoring — Findings (Solo+) — cloud-side detectors surface fleet issues (disk, failed services, slow queries, credential-scanning cross-referenced with fail2ban, container health, TLS expiry, pending updates) as triageable cards, with gated one-click remediation (server-signed preview → human confirm → verb-allowlisted executor; block IP or renew a cert). → guide
• CloudWatch Logs browser — log groups with Top-IPs analysis, IP geolocation, and AbuseIPDB lookups.
• CloudTrail browser — AWS CloudTrail events with region / time / event / user filters.
• IP ban manager — ban IPs via AWS WAF, Security Groups, or NACLs, with an audit trail.
• Keyword server scanning — search file contents across instances.

AI

• Servonaut AI (Solo+) — hosted AI gateway; chat with your fleet with no local API key. The model can tail logs, run commands (with confirmation), and triage incidents over the relay — credentials and SSH keys never leave the CLI. Quota inline / servonaut ai quota.
• Bring your own key — OpenAI / Anthropic / Gemini / Ollama keys configured per-provider in Settings → AI Provider (local Ollama needs none). All coexist with Servonaut AI, switchable per-session.
• Built-in AI chat — LLM assistant with tool-calling against your instances (the same MCP tool surface below).
• AI log analysis — analyze logs with OpenAI, Anthropic, Gemini, or Ollama, with cost estimation.

Memory & secrets

• Server memory — persistent per-server cache of OS / runtime / service / web-stack / log / database / container / git / disk facts; optional background fleet auto-scan. → docs
• Memory Sync (Solo+) — end-to-end-encrypted backup of fleet memory to servonaut.dev (X25519 + AES-256-GCM, your passphrase), with drift detection, cross-device history, and optional auto-sync. ☁ Memory Sync in the sidebar.
• Database credential vault (Solo+) — scan a server for the DB credentials its apps already use, store the password in your secret vault under a per-site label, and let the db_* tools resolve it by name — no password in config or agent context. → docs

Agents & automation (MCP)

• MCP server — ~80 tools for Claude Code, Cursor, Windsurf, etc.: instance ops, AWS / Hetzner / OVH lifecycle, S3, log analysis & IP banning, Docker inspection, system-health probes, SSH-key CRUD, memory queries, and an authenticated REST proxy — behind a three-tier guard (readonly / standard / dangerous) with a JSONL audit trail. → details below
• MCP relay — servonaut connect (or TUI autostart) holds a Mercure SSE connection open so agents and team-mates can dispatch tool calls to this machine. Tokens never leave the CLI.
• Servonaut Cloud account — optional servonaut login unlocks config sync across machines and the MCP relay.
• Config sync — client-side-encrypted snapshots of your config.json synced via servonaut.dev, paired with a passphrase you control.

Convenience

• Instance caching — stale-while-revalidate for fast startup.
• Auto-update — startup check + one-click update (servonaut --update).
• Desktop shortcut — servonaut --install-desktop (Linux/macOS).
• Fully configurable — everything in ~/.servonaut/config.json.

Prerequisites

• Python 3.10+
• AWS CLI configured (~/.aws/credentials and ~/.aws/config)
• SSH client (standard on Linux/macOS, OpenSSH on Windows)
• pipx for isolated installation (recommended)

Your AWS credentials need ec2:DescribeInstances and ec2:DescribeRegions permissions. Additional permissions needed for optional features:

Getting Started

servonaut

That's the whole interface. The TUI is the primary and recommended way to use Servonaut — every feature (fleet view, SSH, remote commands, logs, IP banning, AI chat, server memory, provider management) is reachable from the sidebar, with full mouse and keyboard support.

A few flags you may want on day one:

servonaut --update                # Check for updates and upgrade
servonaut --install-desktop       # Create desktop shortcut (Linux/macOS)
servonaut --setup-ovh             # Guided OVHcloud credential setup
servonaut --debug                 # Verbose logging to stderr

Headless & automation: every major feature also has a scriptable CLI (servonaut connect, servonaut memory, servonaut ai, servonaut hetzner, servonaut secrets) for CI runners, cron jobs, and boxes without an interactive session — see the CLI Reference. Wiring up an AI agent instead? Jump to MCP Server for AI Agents.

Keyboard Shortcuts

What You Can Do

The TUI opens to a unified instance list (AWS + OVH + Hetzner + custom servers in one searchable table). The collapsible left sidebar groups everything else by purpose:

Core

• 📋 Instances — search and SSH the unified fleet
• 💻 Custom Servers — add / edit / remove non-AWS servers (DigitalOcean, on-prem, etc.)
• 🔑 SSH Keys — configure default and per-instance keys

Logs & Security

• 📊 CloudWatch — browse AWS log groups with Top IPs analysis, action filter (All/Allowed/Blocked), IP geolocation, AbuseIPDB lookup
• 🔒 IP Ban Manager — ban IPs via WAF, Security Groups, or NACLs
• 🔍 CloudTrail — audit AWS API activity with filters

Tools

• 🧠 Fleet Memory — scan / refresh / inspect the AI-queryable fact cache, with an optional scheduled background auto-scan (bulk scans run in the background and survive leaving the panel)
• ☁ Memory Sync — encrypted backup of fleet memory across devices (Solo+)
• 🛡 Findings — proactive-monitoring inbox: scan, review, and triage server-detected issues fleet-wide (Solo+; Free shows an upgrade card)
• 🔄 Sync Config — encrypted config snapshots (Solo+)
• 🔧 Settings — configuration, scan rules, AI provider, AbuseIPDB key

OVH (visible when configured)

• ⚙ Manage — table of OVH instances with state-aware lifecycle toolbar (Create / Start / Stop / Reboot / Delete)
• 🔑 SSH Keys — project-level SSH key registry (the one the create wizard injects from)
• DNS Zones · IP Management · Block Storage · Billing

Hetzner (visible when configured)

• ⚙ Manage — table of Hetzner servers with full lifecycle toolbar (Create / Power on / Shutdown / Reboot / Delete)
• 🔑 SSH Keys — Hetzner Cloud project SSH key registry

Account

• Login · Teams · Bug Reports

Server Actions — clicking any instance row opens a per-instance dashboard: the detail pane shows the server's identity, a memory snapshot, and an opt-in live resource monitor (L), and the action rail covers:

• Browse Files (inline) · Run Command · SSH Connect · SCP Transfer
• View Scan Results · View Logs (tail -f) · AI Analysis · Findings (F)
• Ban IP · Manage/Verify SSH Ref

The SSH Ref editor pairs with a Bitwarden vault — pick an SSH key from a list instead of pasting a UUID, or import keys straight from ~/.ssh (passphrase-protected included), so a machine with no local keys can still connect. The TUI, CLI, and MCP agents all resolve the key from your vault at connect time (Solo+). → docs

Command history persists across sessions — Ctrl+R to search history and saved commands, Ctrl+S to save favorites.

Instance Caching

Configuration

All configuration lives in ~/.servonaut/config.json, created automatically on first run.

See Configuration Guide for the full reference including connection profiles, custom servers, scan rules, and match conditions.

SSH keepalives: all connections send keepalives by default so long or idle agent-driven sessions don't get dropped. Tune globally via the ssh block in config.json (server_alive_interval, server_alive_count_max, tcp_keepalive, connect_timeout).

Legacy / special-case SSH hosts: connection profiles and custom servers both accept an extra_ssh_options array that appends arbitrary -o KEY=VALUE flags per host — use it to talk to ancient OpenSSH boxes (HostKeyAlgorithms=+ssh-rsa), override keepalives, or set connect timeouts without weakening your global SSH defaults. See Per-host SSH tuning.

Secrets: API keys in config.json support $ENV_VAR and file:~/.secrets/key syntax so the config file stays secret-free. You can also create ~/.secrets/servonaut.env with KEY=value pairs — loaded automatically on startup.

Optional Dependencies

# MCP server for AI agents
pipx inject servonaut mcp
# or: pip install 'servonaut[mcp]'

# Hetzner Cloud / OVHcloud provider SDKs
pip install 'servonaut[hetzner]'
pip install 'servonaut[ovh]'

# Install everything
pip install 'servonaut[all]'

AI log analysis (OpenAI, Anthropic, Gemini, Ollama) needs no extra install — httpx ships as a base dependency.

MCP Server for AI Agents

This section is for wiring up AI agents (Claude Code, Cursor, Windsurf, …) — not day-to-day interactive use. If you're a human operating your fleet, the TUI above is the recommended interface.

Servonaut includes an integrated MCP server that exposes tools to AI agents like Claude Code:

# Auto-install into a coding agent
servonaut --mcp-install claude     # Claude Code
servonaut --mcp-install cursor     # Cursor
servonaut --mcp-install windsurf   # Windsurf
servonaut --mcp-install opencode   # OpenCode
servonaut --mcp-install vscode     # VS Code Copilot
servonaut --mcp-install all        # All of the above

# Run MCP server manually (stdio transport)
servonaut --mcp

Agent-only / headless install

You don't need the TUI to use Servonaut as an agent toolbox. The MCP server runs fully headless — servonaut --mcp never loads the terminal UI (this is enforced by a regression test), so you can install it on a server or CI box purely as an MCP backend for your coding agent:

pipx install 'servonaut[mcp]'
servonaut --mcp-install claude   # or cursor, windsurf, opencode, vscode, all

Configure credentials and servers the same way as a TUI install ( ~/.servonaut/config.json, $ENV_VAR / file: secret syntax — see Configuration Guide). For Servonaut Cloud features (relay, config sync, hosted AI), servonaut login runs the device-flow sign-in fully headless — approve from a browser on any device. Everything an agent does goes through the same guard levels and is logged to ~/.servonaut/mcp_audit.jsonl.

SSH keys from Bitwarden (no keys on the box). If your instances have a Bitwarden SSH ref saved, the SSH-backed tools (run_command, get_logs, transfer_file, …) resolve the private key from your vault at connect time instead of needing it in ~/.ssh — so an agent on a fresh server or CI box can connect with no local keys at all. Because a headless process can't prompt for your master password, unlock the vault once and export the session into the environment the MCP server (or servonaut connect) runs in:

export BW_SESSION=$(bw unlock --raw)   # unlock once; stays valid until you `bw lock` or the shell exits
servonaut --mcp                         # child inherits BW_SESSION

The key is written to a private, 0600 temporary file only for the duration of each command and deleted immediately after. If the vault is locked or bw isn't installed, the tools fall back to local keys — a working local setup is never affected.

Available tools:

The tool list is filtered to what's actually usable: OVH and Hetzner tools appear only when those providers are configured, the ip_ban_* tools only when at least one IP-ban target is defined, and the *_server_memory* tools only when the memory subsystem is enabled. CloudWatch/CloudTrail and the core instance tools are always available (AWS is the base provider).

• cloudwatch_top_ips parses WAF/ALB structured logs to rank client IPs with allowed/blocked counts — pair it with cloudtrail_lookup_events to corroborate, then ip_ban_set to block via WAF, a security group, or a NACL.
• whoami returns session metadata — the OAuth bearer is never exposed.
• api_request lets an agent make authenticated REST calls against servonaut.dev with automatic 401 refresh and a CLI-side rate limit (30/min). The bearer stays on the CLI.
• mcp_tool_call wraps a JSON-RPC 2.0 tools/call envelope against the hosted MCP at mcp.servonaut.dev — used for premium tools when your plan includes them.
• get_server_memory(id) returns the cached fact snapshot — agents call this BEFORE any SSH round-trip so they answer most OS / runtime / service questions without run_command. Pass format='context_block' to get back a <CONTEXT> envelope for direct prompt injection.

Guard levels: readonly (list/status/introspection only — includes CloudWatch/CloudTrail and ip_ban_list_* queries), standard (read + safe commands + authenticated REST + power management — start / stop / reboot / shutdown + S3 download), dangerous (everything, including create_server / delete_server / transfer_file / ip_ban_set / aws_terminate_instance / aws_run_instances / S3 mutations (s3_create_bucket, s3_delete_bucket, s3_upload_object, s3_delete_object, s3_copy_object, s3_move_object, s3_generate_presigned_url)). Dangerous shell commands (rm -rf, shutdown, reboot, etc.) are always blocked regardless of guard level. Mutating tools carry an explicit "confirm with the user before calling" cue in their descriptions; the top-level MCP instructions document the three-step protocol (summarise → state args → wait for affirmative reply). All operations are logged to ~/.servonaut/mcp_audit.jsonl.

Servonaut Cloud account

Optional — Servonaut works fully offline against your own AWS / OVH credentials. Signing in at servonaut.dev unlocks:

• Config sync — push/pull an encrypted snapshot of your config.json between machines. The passphrase never leaves your client; the server only sees ciphertext. Sidebar entry 🔄 Sync Config opens the snapshot manager directly (Pull Latest / Push New / Restore / Rename / Delete).
• MCP relay — a Mercure SSE channel that lets AI agents and team-mates dispatch MCP tool calls to this machine. While the relay is connected, https://servonaut.dev/account reports your CLI as online, and hosted agents can reach it.
• Memory Sync (Solo+) — encrypted fleet memory backup with drift detection. Open the ☁ Memory Sync sidebar entry, click Unlock Memory Sync, and enter a passphrase. The same screen handles first-time enrolment AND post-restart unlock — your private key is wrapped with the passphrase locally, so the server never sees it. After unlock, click Sync now to push every cached server's memory modules as encrypted envelopes, or flip on auto-sync to drain the queue in the background so the server-side copy (and weekly digests) stay current. Unlock survives restarts: tick Remember on this device to silently re-unlock from your OS keychain on the next launch (re-prompted after 30 days, or Forget on this device to clear it). If you decline, Memory Sync stays dormant until you next open a memory section. Per-feature settings (digest cadence, Mercure push, AI consent) live at the bottom of the Settings panel and are stored on your servonaut.dev account.

Sign in from the TUI's Account / Login screen. After a successful device-flow authentication, the TUI auto-starts an in-process relay listener and the sidebar indicator flips to ● connected.

The listener is tied to the TUI window — closing the TUI drops the connection after ~60 s. For always-on reachability (CI runners, headless boxes), use servonaut connect --bg instead; the CLI and TUI cooperate over ~/.servonaut/relay.lock so they can't run at the same time. The TUI shows external listener (PID N) when a --bg listener is holding the connection.

servonaut connect                 # Foreground relay (Ctrl+C to stop)
servonaut connect --bg            # Detach; writes ~/.servonaut/relay.pid
servonaut connect --status        # Local + backend view with divergence warning
servonaut connect --stop          # SIGTERM the background listener
servonaut connect --reconnect     # Heal a stale SSE socket (stop+start)
servonaut connect --force-bg      # Take over from a TUI's in-process listener

See CLI Reference → servonaut connect for full flags.

Tokens are stored at ~/.servonaut/auth.json with mode 0600, written atomically via tmp + os.replace(). If an older build left the file world-readable, the next run auto-fixes it.

Proactive monitoring — Findings (Solo+)

Server-side detectors scan your fleet for problems and post each one as a triageable finding card: severity, description, evidence, and remediation options. Detection runs in the Servonaut cloud; this client only runs read-only probes over your relay connection and renders the results — nothing is analysed or decided locally.

• Fleet inbox / per-instance — 🛡 Findings in the sidebar for the whole fleet, F on any server for that instance.
• Scan now (s) — dispatches read-only probes over the relay (servonaut connect or the TUI autostart must be running); inapplicable detectors report the reason instead of failing silently.
• Triage — acknowledge (a), resolve (r), or suppress (x); status syncs server-side.
• Gated one-click remediation — automatable fixes run only on an explicit click, from a server-signed preview, through a verb-allowlisted executor (typed confirmation for state changes, dry-run first, fully audited). Fixes today: block a source IP (AWS WAF / Security Group / NACL, or the box's own firewall — nftables / ufw / firewalld) and renew a TLS certificate.

→ Full guide: docs/proactive-monitoring.md — detectors, the detection/probe/remediation model, and the safety & privacy design. Included with Solo and Teams (monitored-instance allowance); Free shows an upgrade card.

Secrets management (Solo+)

Centralise SSH keys + other named secrets behind a pluggable provider backend. Day-to-day this is invisible: once configured, SSH key resolution checks your provider automatically on every connect — you keep clicking SSH in the TUI and it just works. The commands below are one-time setup. MVP supports two backends:

• LocalProvider — keys live in ~/.servonaut/secrets.json (mode 0600, atomic write, same trust model as auth.json). Always available on Solo and Teams plans.
• BitwardenProvider (bws) — keys live in your team's Bitwarden Secrets Manager project. Team admin configures the project from https://servonaut.dev/account/teams/<slug>/secrets; CLI fetches the metadata and reads/writes through the local bws binary using your own access token. The token never leaves your machine — servonaut.dev only stores the project ID and the name of the env var holding the token.

To use Bitwarden as your team's backend:

# 1. Install the bws CLI (one-time)
servonaut secrets install bws         # macOS: brew · Linux: cargo
# Windows / other → prints upstream install URL.

# 2. Mint a BWS access token (https://bitwarden.com/help/personal-access-tokens/)
#    and export it
export BWS_ACCESS_TOKEN=<your-token>

# 3. Verify wiring
servonaut secrets status              # shows plan, entitlement, active provider

# That's it — SSH key resolution now checks Bitwarden first, ~/.ssh as
# fallback. Push a key into BWS with `bws secret create`:
bws secret create "$(basename ~/.ssh/prod-server)" \
                  "$(cat ~/.ssh/prod-server)" \
                  --project-id <project-uuid-from-status>

Key resolution order on every SSH connect:

1. Active provider (Bitwarden, if configured) looked up by key name.
2. ~/.ssh discovery (existing patterns + fuzzy match).
3. The path stored in config.json::instance_keys[<id>] or config.default_key.

Free-tier users get the legacy ~/.ssh-only flow with zero behaviour change. Provider-supplied keys land in ~/.servonaut/keys/<name> at mode 0600.

Threat-model + design notes are pinned in the codebase via inline docstrings on services/secret_provider.py, services/bitwarden_provider.py, and services/secret_provider_resolver.py.

Database credential vault

The same secret store also backs a database credential vault: scan a server for the DB credentials its apps already use — .env / DATABASE_URL (including DATABASE_URL_PROD / _STAGING variants), wp-config.php, configuration.php, Magento env.php, and docker-compose environment: blocks, with a read-only sudo -n fallback so root-owned files and containerized stacks are covered. Store the password under a per-site label and the db_processlist / db_top_queries tools resolve it by name — the password never lands in your config or in an AI agent's context. A Secrets → DB coverage view lists which instances are covered per site, with in-place label and remove. Full docs

Development

# Run directly (primary dev workflow)
PYTHONPATH=src python3 -m servonaut.main

# Run with debug logging
PYTHONPATH=src python3 -m servonaut.main --debug

# Install editable
pip install -e .

# Update pipx installation after changes
pipx install . --force

# Run tests
pip install -e ".[test]"
pytest

See Architecture for codebase structure and design patterns.

Troubleshooting

See Troubleshooting Guide for help with SSH connections, bastion hosts, key management, and AWS credentials.

Runtime Files

All runtime files are under ~/.servonaut/:

Logging

Logs are always written to ~/.servonaut/logs/servonaut.log. Use --debug for verbose stderr output.

When SSH fails, the terminal window stays open showing the error and exit code.

Listed on

Also published to the official MCP Registry as dev.servonaut/servonaut.

License

This project is licensed under the MIT License — see the LICENSE file for details.