The open-source Model Context Protocol (MCP) server for VPS and Linux infrastructure. Turn Claude, Cursor, and agentic LLMs into safe Linux administrators with 45 deterministic tools, disposable Docker migration rehearsals, zero-downtime GitLab CI/CD, and Telegram triage.
Select a scenario below to inspect the MCP tool execution, parameters, and structured JSON-RPC responses.
A dedicated full-stack console built for high-level operations. Complete with real-time SSE streaming, LaTeX math formula rendering, persistent sidebar chats, and live memory inspection.
Renders probability distributions, Bayes' Theorem, algorithmic complexities, and financial statistics flawlessly in chat threads.
Effortlessly switch between system maintenance chats, migration reviews, and secretary briefings with full conversational memory.
Inspect verified architectural decisions, operator preferences, and procedures stored in the structured agent memory store.
Engineered from the ground up to eliminate the risks of autonomous AI agents while multiplying their capabilities.
Communicates over standard SSH using strict sudoers allowlists and rigid command templates. Eliminates arbitrary bash injection vulnerabilities. Every mutation provides a detailed dry-run preview.
Never execute migrations on live production databases. Spawns isolated, disposable Docker containers to test complex Odoo schema upgrades, computes column-level diffs, and automatically purges test copies.
Inspect branches, commits, merge requests, and tail failed CI pipeline logs with automatic token/secret redacting. Trigger verified rolling deployments with instant rollback state pointers.
Equips the AI assistant with dense vector search (Sentence Transformers + FAISS) over technical engineering books and architecture specs. Returns verbatim text snippets with exact page citations and scores.
Connects directly to your personal Telegram account via Telethon MTProto (no generic bot token limits). Scans unread client messages, drafts replies, and waits for your confirmation before sending.
Cross-client knowledge store preserving facts, architectural decisions, operator preferences, and recovery runbooks with confidence scoring, provenance tracking, and expiration policies.
Search and inspect all 45 tools exposed over MCP stdio protocol.
Unconstrained terminal access allows autonomous agents to hallucinate destructive commands or succumb to indirect prompt injections. VPS-MCP creates a zero-trust safety boundary.
rm -rf / or exfiltrating private SSH keys.
confirm=false): Mutating actions generate a detailed preview first. Real execution requires operator approval.
get_audit_log.
VPS-MCP adheres to the standard Model Context Protocol. Works out of the box with Claude Desktop, Cursor, VS Code, and custom agent harnesses.
// Claude Desktop Configuration (~/Library/Application Support/Claude/claude_desktop_config.json)
{
"mcpServers": {
"steward": {
"command": "uvx",
"args": [
"steward-mcp",
"serve",
"--config",
"/absolute/path/to/vps-mcp.toml"
]
}
}
}
uvx steward-mcp serve boots directly into inspection fallback mode, exposing the consolidated compact catalog (or all 45 granular tools with --mode full) for exploration without live credentials.
Clear answers on Model Context Protocol security boundaries, migration rehearsals, AI clients, and deployment architecture.
rm -rf, unindexed SQL updates, or leaked SSH keys can destroy production. VPS-MCP introduces an immutable governance layer: dangerous destructive commands are strictly disallowed, every state modification requires dry-run previews, automated post-action verification checks run automatically, credentials and tokens are redacted, and every call is written to an append-only audit log.