Servagent

Servagent is an MCP (Model Context Protocol) server that enables a remote AI to take full control of a Linux server: command execution, file management, service administration, and more.

Transports

The server exposes two MCP transports simultaneously:

Related MCP server: mcp-remote-agent

Features

Each tool is annotated with MCP ToolAnnotations (read-only, destructive, idempotent) to guide AI clients. Server instructions include anti-loop rules, error handling, and workflow guidelines. Redundant tools (ls, cp, mv, rm, ps, kill, etc.) have been removed — execute_command covers all these use cases.

Prerequisites

• Linux (Ubuntu/Debian, RHEL/CentOS, etc.)

• Python >= 3.10

• Root access for installation as a service

One-liner Installation

Install directly on your server with a single command:

# Simple HTTP installation (by IP)
curl -sSfL https://raw.githubusercontent.com/servagent/servagent/main/install-remote.sh | sudo bash

# HTTPS installation with Let's Encrypt
curl -sSfL https://raw.githubusercontent.com/servagent/servagent/main/install-remote.sh | sudo bash -s -- your-domain.com

# Installation with full sudo privileges + HTTPS
curl -sSfL https://raw.githubusercontent.com/servagent/servagent/main/install-remote.sh | sudo bash -s -- --full-access your-domain.com

# Install a specific version
curl -sSfL https://raw.githubusercontent.com/servagent/servagent/main/install-remote.sh | sudo bash -s -- --version v0.2.0

The script automatically downloads the latest release (or the main branch if no releases exist), extracts the archive, and runs the installation. The -y flag is injected automatically so that interactive prompts are skipped (since stdin is not a terminal in a curl | bash pipeline).

Installation from a Git Clone

If you prefer to clone the repository manually:

# 1. Clone the repository
git clone https://github.com/Servagent/servagent.git
cd servagent

# 2a. Simple HTTP installation (by IP)
sudo bash install.sh

# 2b. OR direct HTTPS installation with Let's Encrypt
sudo bash install.sh your-domain.com

# 2c. OR installation with full sudo privileges (no interactive prompt)
sudo bash install.sh --full-access your-domain.com

The script automatically:

• Creates a servagent system user

• Installs in /opt/servagent with a virtualenv

• Generates an API key (displayed once — save it)

• Creates and enables the systemd service

• If a domain is provided: Let's Encrypt certificate, HTTPS on port 443, auto-renewal

• Interactive prompt to grant full sudo privileges (or --full-access to automate)

# Check status
servagent status

# View logs
sudo journalctl -u servagent -f

The server starts automatically at the end of installation:

• Without domain: http://<server-ip>:8765/mcp (Streamable HTTP) or /sse (SSE)

• With domain: https://your-domain.com/mcp (Streamable HTTP) or /sse (SSE)

Manual Installation (Development)

# Create a virtualenv
python3 -m venv .venv
source .venv/bin/activate

# Install the project
pip install -e .

# Configure
cp .env.example .env
# Edit .env to set SERVAGENT_API_KEY

# Run
servagent

# Available commands
servagent --help
servagent --version

Status

Check the service status and current configuration:

servagent status

Displays: systemd service state (active/inactive/failed), PID, uptime, and configuration summary (port, API key, OAuth, TLS, enabled tools). API keys and OAuth secrets are masked (only the last 6 characters are shown). If the .env file is not readable, the command automatically escalates with sudo.

API Key Management

Generate and manage the API key via the CLI:

servagent apikey setup       # Generate an API key and write it to .env
servagent apikey renew       # Regenerate the API key (invalidates the current one)
servagent apikey remove      # Comment out the API key in .env

The key is displayed in full only at generation time (setup / renew). Use servagent status to verify it is configured (masked).

Uninstallation

To completely remove Servagent from the server:

servagent uninstall              # Interactive (confirmation required)
servagent uninstall -y           # Non-interactive (no confirmation)
servagent uninstall --keep-certs # Keep Let's Encrypt certificates

The script automatically removes:

• The systemd service and certbot renewal timer

• The Nginx configuration (if applicable)

• The sudoers file (/etc/sudoers.d/servagent)

• The application directory (/opt/servagent: virtualenv, .env, sources)

• The servagent system user

• Let's Encrypt certificates (unless --keep-certs is used)

Note: System packages (certbot, python3, nginx) are not removed as they may be used by other services.

Update

Update to the latest version:

servagent update             # Update from the current branch
servagent update develop     # Update from a specific branch
servagent update --force     # Force reinstallation even if already up to date

The script automatically:

1. Runs git pull to fetch the latest changes

2. Copies sources to /opt/servagent/

3. Reinstalls the package in editable mode (pip install -e) in the virtualenv

4. Restarts the service

5. Verifies the service is running correctly

If no changes are detected, the script stops without restarting the service. In case of issues, a rollback command is displayed at the end of execution.

Configuration

All options are configurable via environment variables or .env file:

HTTPS with Let's Encrypt

TLS is built directly into install.sh. Simply pass the domain as an argument (see One-liner Installation).

Prerequisites: The domain must point to the server's IP and port 80 must be open for Let's Encrypt's HTTP-01 challenge.

To enable HTTPS on a server already installed with HTTP:

sudo bash setup-tls.sh your-domain.com

Alternative: Nginx Reverse Proxy

If you prefer to use Nginx (useful when other web services run on the same server):

# 1. Install Nginx and Certbot
sudo apt install nginx certbot python3-certbot-nginx

# 2. Obtain a Let's Encrypt certificate
sudo certbot --nginx -d your-domain.com

# 3. Copy the Nginx configuration
sudo cp nginx.conf.example /etc/nginx/sites-available/servagent
# Edit the file to replace 'your-domain.com' with your domain

# 4. Enable and reload
sudo ln -s /etc/nginx/sites-available/servagent /etc/nginx/sites-enabled/
sudo nginx -t && sudo systemctl reload nginx

The MCP endpoint will then be accessible over HTTPS: https://your-domain.com/mcp

Note: Behind a reverse proxy, set SERVAGENT_HOST=127.0.0.1 so the server automatically disables the MCP SDK's DNS-rebinding protection (Nginx already handles Host header validation).

Connecting from an MCP Client

Claude Code

Claude Code supporte nativement les serveurs MCP distants via Streamable HTTP. Ajoutez cette configuration dans votre fichier de settings Claude Code (.mcp.json, project settings, ou via claude mcp add) :

{
  "mcpServers": {
    "servagent": {
      "type": "streamable-http",
      "url": "https://your-domain.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Claude Desktop

Claude Desktop ne supporte que les serveurs MCP locaux via stdio. Pour se connecter à un serveur distant, utilisez mcp-remote comme pont. Ajoutez ceci dans votre claude_desktop_config.json :

{
  "mcpServers": {
    "servagent": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://your-domain.com/mcp",
        "--header",
        "Authorization: Bearer YOUR_API_KEY"
      ]
    }
  }
}

mcp-remote crée un serveur stdio local qui relaie les requêtes vers le serveur HTTP distant. Nécessite Node.js installé.

Autres clients (LM Studio, etc.)

Les clients qui supportent nativement Streamable HTTP peuvent utiliser cette configuration :

{
  "mcpServers": {
    "servagent": {
      "url": "https://your-domain.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

SSE (Legacy Clients)

{
  "mcpServers": {
    "servagent": {
      "url": "https://your-domain.com/sse",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Testing with curl

# Health check (the MCP endpoint responds to POST requests)
curl -X POST https://your-domain.com/mcp \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

File Upload

The POST /upload endpoint allows sending files to the remote server via multipart/form-data. It is protected by the same Bearer token as the MCP endpoints.

# Send a file to the server
curl -X POST https://your-domain.com/upload \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@my-file.tar.gz" \
  -F "path=/opt/app/my-file.tar.gz" \
  -F "create_dirs=true"

Form fields:

• file (required): the file to send

• path (required): destination path on the remote server

• create_dirs (optional, default true): create parent directories if needed

The maximum size is configurable via SERVAGENT_UPLOAD_MAX_SIZE (default: 100 MB).

Authentication

The server supports two authentication mechanisms that coexist:

Both mechanisms work in parallel. The Bearer token gives direct access to all endpoints. OAuth enables a standard registration and authorization flow.

The /.well-known/ endpoints (OAuth discovery, RFC 8414 / RFC 9728) are always accessible without authentication.

OAuth 2.0 (Streamable HTTP)

In addition to simple Bearer token authentication, the server supports OAuth 2.0 for the /mcp endpoint. This allows compatible MCP applications to connect via the standard OAuth protocol (authorization code + PKCE, RFC 7636).

Enabling OAuth

# Auto-detects the issuer URL from TLS cert domain or server IP
servagent oauth setup

# Explicit issuer URL (if auto-detection fails)
servagent oauth setup --issuer-url https://your-domain.com/mcp

This generates CLIENT_ID and CLIENT_SECRET, and writes all three SERVAGENT_OAUTH_* variables to .env:

# In .env — the URL MUST include the /mcp path
SERVAGENT_OAUTH_ISSUER_URL=https://your-domain.com/mcp

# Operator credentials (dual purpose: static OAuth client + /mcp/register protection)
SERVAGENT_OAUTH_CLIENT_ID=servagent-xxxxxxxxxxxxxxxx
SERVAGENT_OAUTH_CLIENT_SECRET=a-strong-randomly-generated-secret

Renewing / Removing OAuth

# Regenerate credentials (invalidates all existing sessions)
servagent oauth renew

# Disable OAuth entirely (comments out vars in .env, removes database)
servagent oauth remove

# Disable OAuth but keep the database file
servagent oauth remove --keep-db

When OAuth is enabled:

• /mcp is protected by OAuth (the MCP SDK handles tokens) or by the Bearer token (API_KEY)

• /mcp/register is protected by HTTP Basic Auth with CLIENT_ID:CLIENT_SECRET

• /sse, /messages/, /upload remain protected by the simple Bearer token (API_KEY)

• /.well-known/ is always accessible without authentication (OAuth discovery)

Authentication Matrix

Dual Purpose of CLIENT_ID / CLIENT_SECRET

The operator credentials (CLIENT_ID / CLIENT_SECRET) serve two purposes:

1. Static OAuth client: At startup, the server pre-registers these credentials as a valid OAuth client in the SQLite database (via ensure_static_client()). This allows interfaces like Claude.ai and ChatGPT to use the CLIENT_ID/SECRET directly in the OAuth flow (authorize → token) without calling /mcp/register. Redirect URIs for known platforms (Claude.ai, ChatGPT) are pre-configured automatically.

2. /mcp/register protection: The same credentials protect the dynamic registration endpoint via HTTP Basic Auth. Programmatic clients (scripts, SDKs) can register to obtain their own client_id/client_secret.

Both modes coexist: the static client works for UIs (Claude.ai, ChatGPT, etc.), and dynamic registration works for scripts and SDKs.

Connecting from Claude.ai

1. On Claude.ai, go to Settings → Connectors → Add a custom connector

2. Fill in the fields:

   • Name: servagent (or a name of your choice)

   • Remote MCP server URL: https://your-domain.com/mcp

   • OAuth Client ID: the value of SERVAGENT_OAUTH_CLIENT_ID

   • OAuth Client Secret: the value of SERVAGENT_OAUTH_CLIENT_SECRET

3. Confirm — Claude.ai automatically performs the OAuth flow (discovery → authorization → token)

Connecting from ChatGPT

1. On ChatGPT, go to Settings → Connectors → Add a custom connector

2. Fill in the fields:

   • Name: servagent

   • MCP server URL: https://your-domain.com/mcp

   • Client ID: the value of SERVAGENT_OAUTH_CLIENT_ID

   • Client Secret: the value of SERVAGENT_OAUTH_CLIENT_SECRET

3. Confirm — ChatGPT performs the OAuth flow with its redirect URI (https://chatgpt.com/connector_platform_oauth_redirect)

OAuth Endpoints (under /mcp)

Note: The /.well-known/ discovery URLs are accessible both at the domain root level and under /mcp. 307 redirects at the root level forward to the /mcp sub-app to ensure compatibility with all MCP clients.

OAuth Flow (Dynamic Registration)

For programmatic clients using dynamic registration:

# 1. Register a client (HTTP Basic Auth with CLIENT_ID:CLIENT_SECRET)
curl -X POST https://your-domain.com/mcp/register \
  -u "MY_CLIENT_ID:MY_CLIENT_SECRET" \
  -H "Content-Type: application/json" \
  -d '{
    "redirect_uris": ["http://localhost:3000/callback"],
    "client_name": "My App",
    "grant_types": ["authorization_code", "refresh_token"],
    "response_types": ["code"],
    "token_endpoint_auth_method": "client_secret_post"
  }'
# → returns dynamic client_id + client_secret

Then the standard OAuth flow proceeds normally:

1. The client discovers endpoints via GET /.well-known/oauth-authorization-server/mcp (or /mcp/.well-known/oauth-authorization-server)

2. The client obtains an authorization code via /mcp/authorize (with PKCE)

3. The client exchanges the code for an access token via POST /mcp/token

4. The client uses the access token to call /mcp

The server uses an auto-approve model: any registered client (static or dynamic) is considered authorized.

Access tokens expire after 1 hour. Refresh tokens last 30 days with automatic rotation.

Storage

OAuth clients and tokens are persisted in a SQLite database (default: ~/.servagent/oauth.db). Data survives server restarts. The static client is re-registered (upsert) on each startup.

Skills

Skills allow you to enrich the context sent to the LLM with information specific to your server: hosted domains, SMTP credentials, available services, etc. Each skill is a directory containing a SKILL.md file whose content is injected into the MCP instructions.

The skills/ directory is located at the project root (in dev) or in /opt/servagent/skills/ (in production). Its content is git-ignored (.gitignore) as it may contain sensitive information specific to each server.

Structure

skills/
├── .gitkeep
├── webserver/
│   └── SKILL.md
├── smtp/
│   └── SKILL.md
└── docker/
    └── SKILL.md

Example: skills/webserver/SKILL.md

# webserver
Domain: myserver.com (points to this server)
Web root: /var/www/myserver.com
Nginx config: /etc/nginx/sites-available/myserver.com
SSL: Let's Encrypt, auto-renew via certbot timer

Example: skills/smtp/SKILL.md

# smtp
This server can send emails via SMTP.
- Host: smtp.gmail.com
- Port: 587
- User: bot@myserver.com
- Password: xxxx-xxxx-xxxx
- Use: `msmtp` or `swaks` CLI (already installed)

The content of each SKILL.md is injected as-is into the MCP instructions under a ## Skills section. If the file starts with a markdown heading (#), it is used as-is. Otherwise, a ### directory_name heading is added automatically.

Security

WARNING: This server gives full control over the host machine. Secure it properly.

• Always set SERVAGENT_API_KEY in production

• Always use TLS (setup-tls.sh or Nginx) in production

• Restrict port access via firewall (ufw, iptables)

• The service runs under a dedicated user (servagent)

• By default, the user has no sudo privileges (NoNewPrivileges=true)

• The --full-access option (or the interactive prompt) grants sudo NOPASSWD: ALL via /etc/sudoers.d/servagent and disables NoNewPrivileges

Project Structure

servagent/
  src/servagent/
    __init__.py        # Version
    cli.py             # CLI entry point (click subcommands: run, status, uninstall, update, apikey, oauth)
    config.py          # Configuration (pydantic-settings)
    auth.py            # Authentication middleware (Bearer + Basic Auth + OAuth)
    oauth_provider.py  # OAuth 2.0 provider with SQLite storage + static client
    tools.py           # All MCP tools
    server.py          # Server module, Starlette app + MCP (Streamable HTTP + SSE + .well-known)
  skills/              # Skills directory (content in .gitignore)
    .gitkeep
  pyproject.toml                  # Metadata and dependencies
  install.sh                      # Linux installation script (from git clone)
  install-remote.sh               # One-liner installation script (curl | bash)
  uninstall.sh                    # Complete uninstallation script
  setup-tls.sh                    # HTTPS setup with Let's Encrypt
  generate-oauth-credentials.sh   # CLIENT_ID / CLIENT_SECRET generator
  nginx.conf.example              # Nginx configuration (optional)
  .env.example                    # Configuration template

Learn More

MCP (Model Context Protocol)

• modelcontextprotocol.io — Official documentation, guides, and tutorials

• GitHub repository — Specification and links to documentation

• Specification (2025-11-25) — Direct access to the dated specification

Skills

• agentskills.io — Agent Skills documentation

• GitHub repository — Agent Skills specification and examples

License

MIT