MCP Task Queue — EC2 + SSE

Pure Python MCP setup so Cursor on your laptop works on tasks stored in a Postgres DB on EC2. Cursor talks MCP over SSE/HTTPS to the EC2 MCP server; it never touches Postgres directly.

Architecture

Jira Issue
    ↓
n8n Workflow (triggered by Jira)
    ↓ HTTP POST /ingest
EC2: Ingest API (FastAPI) — validates, normalizes, writes to Postgres
    ↓
EC2: Postgres — tasks table (pending, in_progress, completed, failed)
    ↑
EC2: MCP Server (SSE) — tools operate on tasks, exposes /mcp/sse
    ↑ SSE/HTTPS
Laptop: Cursor — configured to connect to EC2 MCP server

Related MCP server: opencode-mcp

Project layout

app/
  __init__.py
  config.py       # DATABASE_URL, INGEST_TOKEN
  database.py     # async Postgres task CRUD
  ingest_api.py   # FastAPI routes: /health, POST /ingest
  mcp_server.py   # FastMCP with tools (list_tasks, get_task, ...)
  mcp_sse.py      # SSE transport wiring

ingest_app.py     # Entry: uvicorn ingest_app:app
mcp_sse_app.py    # Entry: uvicorn mcp_sse_app:app
init_db.py        # Create tasks table
env.example       # Template for .env

Install

python -m venv .venv
source .venv/bin/activate   # Windows: .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp env.example .env
# Edit .env: DATABASE_URL, INGEST_TOKEN

Database migrations

Create the tasks table once:

python init_db.py

Schema: id, source, title, instructions, acceptance_criteria (JSONB), file_hints (JSONB), meta (JSONB), status, timestamps (created_at, started_at, completed_at, failed_at, updated_at), completion_note, failure_reason, previous_status.

Run the services

Ingest API (n8n webhook)

uvicorn ingest_app:app --host 0.0.0.0 --port 8787

• GET /health — health check

• POST /ingest — requires header X-Ingest-Token: <INGEST_TOKEN>

MCP SSE server (Cursor endpoint)

uvicorn mcp_sse_app:app --host 0.0.0.0 --port 8080

• GET /health — health check

• GET /mcp/sse — MCP SSE endpoint (Cursor connects here)

• POST /mcp/messages/ — MCP message endpoint (used by client after SSE)

Configure Cursor (client)

Configure Cursor on your laptop to talk to the EC2 MCP server over SSE. Copy mcp.json.example to .cursor/mcp.json or add to global MCP settings:

{
  "mcpServers": {
    "ec2-tasks": {
      "url": "https://<YOUR_DOMAIN_OR_IP>/mcp/sse",
      "env": {}
    }
  }
}

Replace <YOUR_DOMAIN_OR_IP> with your EC2 hostname or IP. If using a non-standard port, include it: https://example.com:8080/mcp/sse. For local dev: http://localhost:8080/mcp/sse.

Important: Cursor connects to the EC2 MCP server over SSE/HTTPS. EC2 talks to Postgres; Cursor never touches the database.

Rules & work prompt

Edit rules.md in the project root to add rules and instructions the MCP server will include in work prompts. This content is combined with each task and passed to the agent.

• MCP prompt task_work_prompt(task_id) — Returns the full prompt (rules + task) to pass to the agent.

• Tool get_work_prompt(task_id) — Same content as JSON for programmatic use.

MCP tools

n8n HTTP Request (Jira → ingest)

• Method: POST

• URL: https://<YOUR_EC2>/ingest (or via ngrok / load balancer)

• Headers: Content-Type: application/json, X-Ingest-Token: <INGEST_TOKEN>

• Body: JSON with at least id, instructions; optional source, title, acceptance_criteria, file_hints, meta

EC2 deployment notes

1. Environment: Set DATABASE_URL and INGEST_TOKEN (e.g. via .env or systemd env).

2. Processes:

   • Ingest: uvicorn ingest_app:app --host 0.0.0.0 --port 8787

   • MCP SSE: uvicorn mcp_sse_app:app --host 0.0.0.0 --port 8080

3. Reverse proxy (Nginx):

   • Proxy /ingest and /health to ingest on 8787.

   • Proxy /mcp/sse and /mcp/messages/ to MCP on 8080.

   • Use HTTPS and keep MCP paths under the same origin.

4. ngrok:

   • Expose both ports or put a reverse proxy in front and expose one.

   • Example: ngrok http 8080 for MCP, ngrok http 8787 for ingest (separate tunnels).

Ingest behavior

• New task → insert as pending.

• Existing, status pending or in_progress → return duplicate error (no overwrite).

• Existing, status completed or failed → upsert as new pending version, preserve previous_status.