MCP Server

Pocket Alert speaks the Model Context Protocol, so an AI agent can send you push notifications, read your message history and check delivery stats as ordinary tool calls.

The practical version: an agent starts a long job, finishes at some unpredictable point, and pings your phone instead of waiting for you to check the terminal.

https://mcp.pocketalert.app/mcp

Connecting

Authentication is a bearer token — your API key from API Keys, sent as Authorization: Bearer <token>.

Claude Desktop / Claude Code

{
  "mcpServers": {
    "pocketalert": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp.pocketalert.app/mcp",
        "--header",
        "Authorization: Bearer ${POCKETALERT_API_TOKEN}"
      ],
      "env": {
        "POCKETALERT_API_TOKEN": "your-api-key"
      }
    }
  }
}

Native HTTP transport

{
  "mcpServers": {
    "pocketalert": {
      "type": "http",
      "url": "https://mcp.pocketalert.app/mcp",
      "headers": {
        "Authorization": "Bearer your-api-key"
      }
    }
  }
}

Restart the client afterwards, then ask it to list its tools — pocketalert should be there.

Available tools

Tool	What it does
list_applications	List your applications
get_application	Fetch one application by TID
list_devices	List registered devices
get_device	Fetch one device by TID
list_messages	Read message history
get_message	Fetch a single message by TID
create_message	Send a push notification
create_message_with_attachment	Send a push with an image attached
delete_message	Delete a message
get_message_stats	Delivery statistics

create_message accepts the same level and actions parameters as the REST API, so an agent can raise a critical alert or attach approval buttons to what it sends.

Worth setting up

Long-running jobs. Migration, training run, batch import — the agent notifies you when it lands rather than you polling a terminal.

Human-in-the-loop approvals. Give the agent create_message and have it send two http action buttons before it does anything irreversible. You approve from the lock screen; the agent proceeds.

Reading its own history. An agent that can call list_messages can answer "what broke overnight?" without you forwarding anything.

Scope the key

An MCP server is a capability you hand to a model. Issue a dedicated API key for it rather than reusing the one your production scripts hold — then revoking the agent's access costs you nothing else.

Related

• Priority levels — what level does to the delivered alert
• Action buttons — approval workflows from a notification
• Create Message — the REST equivalent of create_message