🚀 MCP Durable Tasks Backend (io.modelcontextprotocol/tasks)

1. What does the MCP Durable Tasks Backend do?

AI Agents (like Claude, Cursor, and custom LangChain bots) natively time out if a tool takes longer than 60 seconds to execute. This Actor provides a plug-and-play Server-Sent Events (SSE) backend that fully implements the July 2026 io.modelcontextprotocol/tasks extension. It allows your MCP server to instantly offload heavy, long-running workloads (like deep web scraping, video rendering, or massive data analysis) into a durable Supabase-backed async queue, returning a task handle to the AI so it can poll for progress without crashing.

2. Why use this Actor? (The Problem it Solves)

• Bypasses the 60-Second AI Timeout: Stop watching your AI agents crash on TimeoutError. Hand them a task ID and let them check back later.
• Zero-Infrastructure Job Queue: You don't need to spin up Redis, Celery, or RabbitMQ. We handle the PostgreSQL/Supabase state management natively.
• Official Protocol Compliance: 100% compliant with the official ext-tasks specification (tasks/get, tasks/update, tasks/cancel).
• Plug-and-Play Middleware: Connects seamlessly to your existing FastMCP, Express, or standard Model Context Protocol servers.

3. Ideal Use Cases

• Autonomous Researchers: Agents that need to scrape 100+ pages of a domain over several minutes.
• DevOps/Infra Bots: Agents triggering long-running CI/CD pipelines, database migrations, or server provisions.
• Media Generation: Handing off video rendering, heavy TTS generation, or bulk image processing workloads.
• Multi-Agent Orchestration: Delegating complex sub-tasks to child agents and monitoring their percentage completion.

4. Input Configuration

The Actor accepts standard JSON configuration to bind to your Supabase instance and define task metadata:

{
  "supabase_url": "https://your-project.supabase.co",
  "supabase_service_key": "eyJhb...",
  "tenant_id": "00000000-0000-0000-0000-000000000001",
  "port": 8002
}

5. Output Data Format

When queried via the get_task_status tool, the Actor returns fully compliant JSON-RPC payloads mapping directly to the MCP spec:

{
  "taskId": "d59e55a9-7056-4805-9597-3bc34283701b",
  "status": "working",
  "progress": {
    "percent": 33,
    "completed_items": 1,
    "total_items": 3
  },
  "result": null,
  "error": null,
  "ttlMs": 3000
}

6. How to Integrate (Code Example)

Simply point your AI Client (Claude Desktop, Cursor, or custom script) to this Apify Standby Actor's SSE URL:

{
  "mcpServers": {
    "durable-tasks": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/inspector", "https://your-apify-actor-url.apify.tech/sse"]
    }
  }
}

7. Cost & Pricing Strategy

• Pay-per-Event: Billed dynamically per asynchronous task created ($0.005/task). Highly cost-effective for intermittent agent operations.

8. Integrations & Compatibility

• Fully compatible with FastMCP (Python), @modelcontextprotocol/sdk (Node.js), Claude Desktop, and Cursor.
• Built natively for Supabase / PostgREST state storage.

9. Limitations & Support

• Max Task Duration: Bound by your Supabase database retention policies and Apify Standby limits.
• Stateless Reconnection: If an SSE connection drops, the AI agent must re-invoke get_task_status(taskId) to resume polling. The state is durably backed up, so no data is lost during disconnects.
• For custom deployments or enterprise SLA, contact via Apify Issues.