How do I push a task to my own AI agent from my phone?

The Vilix Agent Inbox lets you push a message to your own running agents from anywhere, including your phone. It rides the agent’s next memory read, so an always-on agent picks the work up on its own. Here is exactly how the inbox field, normal vs urgent delivery, and ack_message work.

You push a task to your own running AI agent from your phone with the Vilix Agent Inbox. From any connected client, you send a short message addressed to your agents, and Vilix holds it until that agent next reads its memory. The next time the agent calls get_context (or a search, or get_project), your note is waiting in the inbox field of the response, and the agent acts on it. There is no new session to start and nothing to copy-paste. This post explains the mechanic honestly: how the inbox field works, the difference between normal and urgent delivery, and when the agent calls ack_message.

What the Agent Inbox actually does

The Agent Inbox is a push channel for your own agents. You write a message, choose who it is for, and Vilix queues it. It is delivered inside the inbox field of that agent’s next memory read. It is not a real-time interrupt and it does not wake a sleeping process. It rides the next read. If your agent is idle, the message simply waits until the agent reads memory again, which for an always-on agent is on its next turn.

That distinction matters, so it is worth being precise. Vilix is an MCP-native memory layer. Agents call get_context before replying and save_turn after, and they call read tools like the search functions, recent_messages, and get_project as they work. The Agent Inbox piggybacks on those reads: any of them returns your queued message in the inbox field. So the speed of delivery is the cadence of the agent’s reads, not a separate notification system.

Normal vs urgent, and ack_message

Every inbox message has an urgency, and it changes how persistent the delivery is.

• Normal urgency (the default) is delivered once per chat, then it is done. It is the right choice for a hand-off you only need the agent to see one time, like starting the next task.
• Urgent urgency is redelivered on every read, in every chat, until the receiving agent calls ack_message to acknowledge it, or until the message expires. Use it for something the agent must not miss, like stop work on the migration.

Messages have a time to live. The default is 24 hours and the maximum is 168 hours (7 days), and you can hold up to 20 unexpired messages at a time. Bodies are Markdown, up to 2000 characters, so a hand-off can carry real detail and not just a one-liner.

You decide who gets the message

An inbox message can be scoped three ways, so it reaches exactly the agents you mean:

• Global goes to all of your agents. Any agent of yours that reads memory will see it.
• Scoped to a project or task you own, so only an agent working on that project or task picks it up.
• Scoped to one source or platform, like Claude Code, so only agents running on that tool see it.

A worked example: one line from your phone, picked up on the next read

Say you run an always-on agent that handles your deploys, wired to Telegram through a harness like OpenClaw or Hermes. You are away from your desk, on your phone, when staging goes green. You open Telegram and send one line to your agent: tell my coding agent the staging deploy is green, start the smoke tests.

That message lands in the Agent Inbox, scoped to the agent or project you chose. Minutes later, the agent comes around to its next turn and calls get_context. Your note is there in the inbox field. The agent reads it, kicks off the smoke tests, and reports back, all without you opening a laptop or starting a new session. If you had marked it urgent, it would keep reappearing on every read until the agent called ack_message, which is exactly what you want for something that must not slip through.

This is the same idea as running one always-on agent across WhatsApp, Telegram, and Slack with shared memory, except here you are not chatting with the agent, you are handing it work to pick up on its own. For a sense of how much an always-on agent accumulates over time, see the 12k-memories always-on agent case study.

Which tools can send and receive inbox messages

Any tool connected to Vilix as a custom MCP connector can send to the inbox and read from it. That covers ChatGPT, Claude, Claude Code, Cursor, Codex, Grok, Manus, GitHub Copilot, Windsurf, and Lovable, plus purpose-built agent harnesses like OpenClaw and Hermes, and any custom agent you wire up yourself. Because the inbox rides ordinary memory reads, you do not add a new integration to receive messages. If an agent already reads Vilix memory, it already has an inbox.

The underlying wiring is the same one-time MCP connection that gives every agent persistent memory in the first place. If you have not set that up yet, the guide to adding memory to any agent with the Vilix MCP walks through it, and the runtime-specific posts for OpenClaw and Hermes cover those harnesses in detail.

Setting it up

The Agent Inbox is part of the same memory layer that works across ChatGPT, Claude, Cursor, Codex, and other AI tools. You connect Vilix once by adding api.vilix.ai/mcp as a custom MCP endpoint. From then on your agents call get_context before each reply and save_turn after, and any of their reads will surface inbox messages addressed to them. The MCP memory server setup guide covers the connection step for each supported tool.

Vilix has a free tier to get started. Try Vilix free and push your first task to a running agent in a few minutes.

Frequently asked questions

How do I push a task to my own AI agent from my phone?

Send a message through the Vilix Agent Inbox from any client connected to your account, including a phone-based one. Vilix queues it and delivers it inside the inbox field of your agent’s next memory read, so the agent picks the work up on its own next turn without you starting a new session.

Is the Agent Inbox a real-time push notification?

No. The Agent Inbox is not a real-time interrupt and does not wake an idle process. It rides the agent’s next memory read: the message waits until the agent next calls get_context, a search, recent_messages, or get_project, and then it appears in the inbox field. For an always-on agent that reads memory every turn, that is effectively the next turn.

What is the difference between normal and urgent inbox messages?

A normal message (the default) is delivered once per chat and then it is done. An urgent message is redelivered on every read, in every chat, until the receiving agent calls ack_message to acknowledge it or the message expires. Use urgent only for things the agent must not miss.

How long does an inbox message last?

Each message has a time to live that defaults to 24 hours, with a maximum of 168 hours (7 days). You can have up to 20 unexpired messages at once, and each body is Markdown up to 2000 characters.

Which AI tools can receive an inbox message?

Any agent connected to Vilix as a custom MCP connector, including ChatGPT, Claude, Claude Code, Cursor, Codex, Grok, Manus, GitHub Copilot, Windsurf, Lovable, and harnesses like OpenClaw and Hermes. Because delivery rides ordinary memory reads, no extra integration is needed to receive messages.

Is there a free plan?

Yes. Vilix has a free tier with no time limit, plus a Pro plan at $19.99 per month with a 7-day full-Pro trial. You can get started for free without a credit card.