Apscheduler

MCP (Model Context Protocol) HTTP server implementation for the apscheduler.

Persistence

The scheduler now supports persistent storage using SQLAlchemy jobstore. Scheduled jobs are stored in the database and survive service restarts.

• Development: Uses SQLite database (dev.db)
• Production: Uses PostgreSQL database (telegram_bot)

Configure via environment variables:

• STORAGE_DB: Set to sqlite for development or postgres for production
• PostgreSQL settings: PG_USER, PG_PASSWORD, PG_HOST, PG_PORT

Example / Events flow / scenario (Pizza Friday)

• Context: An agent is connected to this scheduler MCP service and to other MCP services:

  • Pizza ordering MCP: tool order_pizza(type-of-pizza, "time to delivery today")
  • Email MCP: tool send_email(email, message) to invite colleagues
  • Bamboo MCP: tool to retrieve birthdays, e.g. get_birthdays(window) where window can be last_week

• User request to the agent: "Every Friday, schedule a pizza order for delivery at 5:00 PM, and schedule an email at 4:45 PM inviting colleagues to Pizza Friday. Include in the invitation the people who had birthdays last week."

• Agent behavior:

  • Retrieves available MCP service endpoints and their tools
  • Builds an execution plan
  • Calls one of the scheduling tools in this service with the plan in JSON

• Scheduler behavior:

  • Accepts the schedule request and stores the cron triggers
  • When time matches, executes the plan

• LLM-based worker behavior:

  • A worker LLM connected to all the MCP services receives the plan at trigger time and executes each action by calling the referenced MCP tools, honoring each action's condition field to manage ordering and success dependencies.

There are two independent schedules below: (1) a single action to order pizza at 17:00 every Friday; (2) two actions at 16:45 every Friday — first get last week's birthdays, then send the email. The second action is conditioned on the first succeeding. The message shows a common placeholder pattern to include upstream results; actual interpolation depends on the worker/agent that executes the plan.

Schedule 1: Order pizza every Friday at 17:00

Tool call input to schedule_tool_call_by_cron:

{
  "execution_plan": {
    "order_pizza": {
      "mcp-service-endpoint": "http://pizza-mcp:3001/mcp/",
      "mcp-tool-name": "order_pizza",
      "mcp-tool-arguments": {
        "type-of-pizza": "assorted",
        "time to delivery today": "17:00"
      }
    }
  },
  "day_of_week": "fri",
  "hour": "17",
  "minute": "00",
  "timezone": "America/Los_Angeles"
}

Schedule 2: Fetch birthdays then send email every Friday at 16:45

Tool call input to schedule_tool_call_by_cron:

{
  "execution_plan": {
    "fetch_birthdays": {
      "mcp-service-endpoint": "http://bamboo-mcp:3003/mcp/",
      "mcp-tool-name": "get_birthdays",
      "mcp-tool-arguments": {
        "window": "last_week"
      },
      "condition": "executes first"
    },
    "send_email": {
      "mcp-service-endpoint": "http://email-mcp:3002/mcp/",
      "mcp-tool-name": "send_email",
      "mcp-tool-arguments": {
        "email": "[email protected]",
        "message": "Welcome Pizza Friday in the kitchen at 4:45pm! Birthdays: ${fetch_birthdays.result.summary}"
      },
      "condition": "executes only if fetch_birthdays was successful"
    }
  },
  "day_of_week": "fri",
  "hour": "16",
  "minute": "45",
  "timezone": "America/Los_Angeles"
}

[!NOTE]

• Endpoints are examples; use your actual MCP service URLs.
• If you enable the worker strategy, the plan is forwarded to the worker MCP > tool and executed there; ensure that worker supports referencing previous action outputs if you rely on placeholders in later actions.

Installation

Environment Variables / .env File

Set up environment variables or configure the .env file using the .env.example as a template.

Common variables:

• MCP_HOST and MCP_PORT: Configure the HTTP host and port the service binds to and exposes. Examples are provided in .env.example.

Execution strategies

This service supports two strategies for executing an execution_plan:

• sequential (default): Executes each action one-by-one in this service process. Set with EXECUTION_STRATEGY=sequential (default when unset).
• worker: Forwards the whole plan to a worker MCP tool which then executes the actions. Enable with EXECUTION_STRATEGY=worker.

When using the worker strategy, configure both:

• WORKER_ENDPOINT: The worker MCP service endpoint
• WORKER_TOOL_NAME: The worker MCP tool name that accepts the plan (called with { "execution_plan": <plan> })

You can set these via environment variables (see .env.example) or at runtime via MCP tools tagged admin:

• set_worker_endpoint(worker_endpoint) — sets WORKER_ENDPOINT
• set_worker_tool_name(worker_tool_name) — sets WORKER_TOOL_NAME

Execution plan schema

The execution_plan parameter annotation/schema is defined in plan-schema.ann. The format of this file is not fixed; it is used as a description shown to MCP agents so they can read and populate the plan. Minimum required fields for every action are:

• mcp-service-endpoint
• mcp-tool-name

Optional per-action field:

• mcp-tool-arguments

You can override the annotation file path using the PLAN_SCHEMA_ANNOTATION_PATH environment variable (defaults to plan-schema.ann).

Docker

A Dockerfile is provided so you can build and run the service in a container.

Build the image:

docker build -t apscheduler-mcp-server:latest .

Run the container (example maps host port 3000 to container port 3000):

docker run --rm \
  -p 3000:3000 \
  -e MCP_HOST=0.0.0.0 \
  -e MCP_PORT=3000 \
  --env-file .env \
  apscheduler-mcp-server:latest

Note: Ensure MCP_HOST is set to 0.0.0.0 in the container so the HTTP server is reachable from outside the container.

Development

Setup

1. Clone the repository.
2. Install development dependencies:

   uv sync --dev
   

3. Create a .env file from .env.example.

Pre-commit Hooks

This project uses pre-commit to automatically check code style with ruff before each commit.

To set up pre-commit hooks, run:

uv run pre-commit install

This will ensure that ruff checks are run automatically before every commit.

Running the MCP Service

To start the MCP service, run:

uv run --env-file .env src/main.py

Available MCP tools

• list_scheduled_jobs() — Lists all scheduled jobs (new)
• remove_scheduled_job(job_id) — Removes a scheduled job by id
• schedule_tool_call_by_cron(execution_plan, ...) — Cron-style scheduling
• schedule_tool_call_at_interval(execution_plan, ...) — Fixed interval scheduling
• schedule_tool_call_once_at_date(execution_plan, run_date) — One-off scheduling
• set_worker_endpoint(worker_endpoint) — Admin tool to set WORKER_ENDPOINT
• set_worker_tool_name(worker_tool_name) — Admin tool to set WORKER_TOOL_NAME

Running Tests

To run the test suite, use:

uv run --dev pytest

Running Linters

This project uses the ruff tool as a linter.

To check code style and linting issues, run:

uv run ruff check

To automatically fix formatting issues, run:

uv run ruff format