MCPCloud

Self-hosted MCP (Model Context Protocol) gateway. Write any Python function, register it as a skill, and it instantly becomes a tool that Claude Desktop, Claude API, Cursor, or any MCP-compatible client can call.

Website: mcpcloud.dev · Demo: demo.mcpcloud.dev · GitHub: carsor007/mcpcloud

How it works

Claude Desktop / Claude API / any MCP client
            ↓  MCP tool call
        MCPCloud  (this repo)
            ↓  your code runs
    Jira · Slack · Salesforce · anything

Every Python function registered as a skill becomes an MCP tool. No vendor lock-in, no proprietary agent framework — just functions.

Quickstart

Deploy to Cloud

Subscribe on AWS Marketplace for a fully managed deployment on ECS Fargate — no Docker, no ECR setup, no servers to manage. Includes a 7-day free trial.

Run locally with Docker

git clone https://github.com/carsor007/mcpcloud.git
cd mcpcloud
cp .env.example .env
docker compose up

Open http://localhost:8000/ui — the tool browser shows all registered skills.

Connect an MCP client

MCPCloud exposes a config endpoint for every agent type. Use it to generate a ready-to-paste snippet for any MCP client.

Claude Desktop

1. Install Claude Desktop from claude.ai/download

2. Get the config snippet for the skills you want:

curl http://localhost:8000/mcp/jira_ops/config

3. Open (or create) ~/Library/Application Support/Claude/claude_desktop_config.json and add the returned snippet under mcpServers:

{
  "mcpServers": {
    "jira_ops": {
      "url": "http://localhost:8000/mcp/jira_ops",
      "transport": "http"
    },
    "slack_ops": {
      "url": "http://localhost:8000/mcp/slack_ops",
      "transport": "http"
    }
  }
}

4. Restart Claude Desktop. Your skills appear as tools in every conversation.

Cloud deployment: replace http://localhost:8000 with your MCPCloud URL (e.g. https://demo.mcpcloud.dev).

Cursor

Add the same JSON to .cursor/mcp.json in your project root, or globally at ~/.cursor/mcp.json. The URL and transport fields are identical.

Claude API (programmatic)

Pass the MCP server URL when initializing a client session — the endpoint follows the MCP 2025-03-26 spec over HTTP.

Adding a skill

Two ways to add a skill: drop a file into skills/ (loaded on startup, lives in your repo), or use the browser editor (saved to Redis, live immediately, no restart or redeploy).

Option A: filesystem

Drop a .py file into skills/. Any file with a register_all() function is loaded automatically on startup.

# skills/my_tools.py
from registry import SkillResult, get_registry

async def my_skill(input: dict, context: dict) -> SkillResult:
    '''One-line description shown in the UI.'''
    return SkillResult(success=True, output={"result": input.get("text", "")})

def register_all():
    get_registry().register(
        "my_tools",   # agent type  — groups skills in the sidebar
        "my_skill",   # skill name  — shown under the group
        my_skill,
        schema={
            "type": "object",
            "required": ["text"],
            "properties": {
                "text": {"type": "string", "description": "Input text"}
            }
        }
    )

Restart the server. The skill appears in the UI and is immediately callable as an MCP tool.

Option B: browser editor

Open /ui/skills for a Monaco-based editor — no filesystem or restart required. Write an async function whose name matches the skill name:

async def my_skill(input: dict, ctx: dict) -> dict:
    """One-line description shown in the UI."""
    return {"success": True, "result": input.get("text", "")}

# Optional: JSON Schema for inputs shown in the tool browser
# SCHEMA = {"type": "object", "properties": {"text": {"type": "string"}}}

Saving POSTs the code to /api/skills, where it's stored in Redis (or in-memory if Redis isn't configured) and registered immediately — the skill is callable as soon as you save, with no restart. This is a separate store from the filesystem skills in skills/, so it survives redeploys of the container image but not a Redis wipe; check code in with Option A if you want it version-controlled.

Included skills

Both work out of the box — real API calls run when credentials are configured, stub data is returned otherwise.

jira_ops

Configure by setting JIRA_URL, JIRA_EMAIL, JIRA_API_TOKEN in .env.

slack_ops

Configure by setting SLACK_WEBHOOK_URL in .env.

Audit log

Every tool call is recorded — caller, tool name, arguments, success/failure, and latency — and available at:

• UI: http://localhost:8000/ui/audit — live table, auto-refreshes every 3s, filterable by agent type/skill name

• API: GET /api/audit?agent_type=jira_ops&skill_name=get_ticket&limit=100

Entries are also emitted as structured JSON log lines, so they flow into CloudWatch Logs (or any log pipeline) for SIEM ingestion. Backed by Redis when REDIS_URL is set (survives restarts, shared across workers); falls back to an in-memory ring buffer otherwise.

Configuration

License

Apache 2.0 — see LICENSE.

Free to self-host. Managed deployment available on AWS Marketplace.