AI Engineering Projects

Remote MCP · Agent Infrastructure

Central AI Agent Hub with MCP for Enterprise-Scale Skill Deployment

1central skill catalog for teams and devices
4data engineering skill packs managed centrally
100%cross-device reuse through one MCP endpoint
MCPModel Context ProtocolTypeScriptNode.jsExpress@modelcontextprotocol/sdkDockerStreamable HTTPSSECursorClaudeZodRemote MCPRender
View on GitHub

Overview

This project solves a core enterprise AI operations problem: teams build useful skills and agent playbooks, but they remain scattered across local machines and become difficult to govern. By hosting those skills behind one MCP endpoint, organizations can distribute agent behavior from a central repository with clear version control, predictable rollout, and simpler onboarding.

Business value comes from standardization and speed. Data engineering teams get reusable guided workflows (data engineering patterns, Databricks secrets practices, workflow builder logic, and Delta table operations) without each engineer reinventing prompts or setup. The same managed catalog can be consumed from multiple clients and devices, reducing drift between environments and improving delivery quality.

The platform model also supports maintainability at scale: new skills are added once, reviewed once, and deployed once. That creates better change tracking, safer updates, and easier auditability of what agent capabilities are available in production at any point in time.

Quick Start

Run a single managed MCP endpoint so data-engineering agent skills stay centralized, versioned, and consistent across teams:

DeployDocker / PaaS: set MCP_TRANSPORT, HOST, MCP_ALLOWED_HOSTS, MCP_AUTH_TOKEN, SKILLS_ROOT — ship one service for the whole org.
ConnectMCP clients: https://<host>/sse (or /mcp) + Authorization: Bearer <token> — same catalog on every device.
GovernVersion SKILL.md packs in git; merge + redeploy to roll updates — one source of truth for agent behavior.

Works across Cursor · Claude Desktop / CLI · any MCP-compatible client over HTTPS (or stdio for local spawning).

Use Cases

Data engineering standards available on every device

Deploy once and consume from multiple laptops and AI clients, so every engineer uses the same approved data engineering skills instead of maintaining local copies.

Reusable Databricks and Delta playbooks

Databricks secrets, workflow builder, and Delta table operation skills remain centrally versioned and reusable across projects, reducing repeated setup and inconsistent implementations.

Governed agent rollout with version control

A central MCP-managed skill repository enables controlled updates, team-wide rollouts, and traceable changes to agent behavior, improving governance in larger organizations.

Faster agent deployment lifecycle

New or updated data engineering skills can be published once and immediately consumed across environments, cutting time-to-adoption for internal AI agents.

Technical delivery architecture

01

SkillsRegistry & MCP server

Walks SKILLS_ROOT recursively for SKILL.md; builds stable ids (POSIX-relative paths); extracts title/description teaser for listings. create-server.ts wires McpServer with registerResource (URI template), registerTool, registerPrompt.

02

HTTP transports

createMcpExpressApp() applies host validation when allowedHosts are set. POST /mcp and POST /sse run stateless StreamableHTTPServerTransport per request—compatible with Cursor’s POST-to-/sse probe. GET /sse opens SSE sessions; POST /messages routes JSON-RPC posts by session id.

03

Auth & compatibility

Bearer middleware accepts Bearer <token> or a bare single token for misconfigured clients; constant-time equality vs MCP_AUTH_TOKEN. Stub POST /register returns JSON when clients attempt OAuth fallback after 401.

04

Ship & observe

Multi-stage Dockerfile (Alpine, non-root user), Compose for local parity, HOSTING_AND_CURSOR.md / DEPLOYMENT.md for Render env vars (MCP_TRANSPORT, HOST, MCP_ALLOWED_HOSTS, MCP_AUTH_TOKEN, SKILLS_ROOT).

Governed skill catalog

data-engineering

Centralized skill guidance for data engineering agent workflows and review patterns.

databricks secrets

Operational practices for Databricks secret handling and safe configuration patterns.

databricks workflow builder

Reusable approach for composing Databricks task graphs and workflow definitions.

delta table operations

Consistent playbook for Delta table lifecycle operations, migrations, and reliability checks.

Tech stack

  • ▸TypeScript / Node.js 20+
  • ▸@modelcontextprotocol/sdk (McpServer, StreamableHTTPServerTransport, SSEServerTransport, StdioServerTransport)
  • ▸Express (createMcpExpressApp)
  • ▸Zod (tool input schemas)
  • ▸Docker & docker-compose
  • ▸Bearer auth + timing-safe compare
  • ▸HTTPS termination at reverse proxy (Render, nginx, etc.)
All AI Engineering ProjectsAll Projects →