Computers and Software

How to Build a 24/7 JARVIS-Style AI Assistant with GPT 6 Astra

I Built My Own JARVIS with GPT 6 Astra (Runs 24/7)

Video: I Built My Own JARVIS with GPT 6 Astra (Runs 24/7)

· based on the channel Damian Malliaros

Key takeaways

  • GPT 6 Astra is the reasoning core; the always-on part comes from running the agent on a cloud server instead of a local machine.
  • Cloudways managed hosting keeps the agent running 24/7 and offers 25% off the first three bills with code DAMIAN.
  • Composio connects the assistant to Gmail, Google Calendar, Slack and thousands of other apps without custom API code.
  • ElevenLabs gives the assistant a real voice, and the whole stack is assembled without writing code.
  • The build keeps the agent on an existing GPT subscription instead of paying per-token API prices.

Building a JARVIS-style assistant with GPT 6 Astra means putting the model in a cloud-hosted agent loop that runs around the clock, wiring it to your apps through Composio, adding an automated email reviewer and giving it a voice with ElevenLabs. In practice the model is only the brain: the always-on behaviour comes from a server that never sleeps, so the assistant keeps working when your laptop is off. The setup described by Damian Malliaros runs on managed cloud hosting — you can start at Cloudways managed AI agents and use code DAMIAN to save 25% on the first three bills.

What a GPT 6 Astra JARVIS actually consists of

A working personal AI assistant is a stack of four separate layers, and confusing them is the main reason most home-built agents die after a week. The first layer is the model — GPT 6 Astra — which handles reasoning, planning and writing. The second is the runtime: a machine that stays online, holds the prompt, schedules the agent and survives reboots. The third is the tool layer, where Composio exposes Gmail, Google Calendar, Slack and thousands of other services as callable actions. The fourth is the output layer: chat, email drafts, morning briefs and, optionally, speech through ElevenLabs.

A useful way to think about it: the model decides what to do, the cloud server decides when, Composio decides how it touches your data, and ElevenLabs decides how it sounds. Each layer can be swapped independently, which matters because model access, hosting prices and integration catalogues change faster than the rest of the stack.

Why the assistant has to live in the cloud

An assistant that runs on your desktop is only available while the desktop is on, which rules out the whole point of a JARVIS: waking up to a brief that was prepared overnight. A cloud server removes that constraint. The agent can poll your calendar at 6 a.m., scan the inbox, draft replies and post a summary to Slack before you touch a keyboard.

Managed hosting is the pragmatic choice here because a self-configured VPS needs patching, monitoring and restart handling, and a crashed process silently means a dead assistant. A managed AI-agent environment handles the operating system, keeps the process alive and gives you a place to store environment variables and API keys. The trade-off is a monthly bill, which is exactly why the discount code matters for the first three invoices.

Deploying the agent on Cloudways step by step

The deployment is a sequence of small, verifiable steps rather than one big configuration. Work through them in order and test after each one.

  1. Create a Cloudways account through the managed AI agents page and apply the code DAMIAN at checkout to get 25% off the first three bills.
  2. Provision a server in the region closest to you; latency matters less for scheduled jobs than for interactive chat, but it still affects voice replies.
  3. Open the application's environment settings and add your model credentials and any integration keys as environment variables, never in the source files.
  4. Deploy the agent application — the video shares the full code so the whole system can be cloned rather than rebuilt.
  5. Set the process to restart automatically and confirm it survives a manual reboot before you connect anything important.
  6. Add a scheduler entry for the morning routine so the brief and the email review fire at a fixed local time.

The reboot test is the one people skip. An agent that works until the first server restart is not a 24/7 assistant, it is a demo.

Configuring GPT 6 Astra and keeping costs down

GPT 6 Astra is configured through a system prompt that defines the assistant's role, its available tools, its tone and its boundaries. The prompt is where you decide whether the agent may send email on its own or only draft it, which calendars it can read, and how it should behave when a tool call fails. Keep the instructions short and specific; long personality essays tend to degrade tool-calling reliability.

On cost, the practical approach shown in the build is to keep the agent running on an existing GPT subscription instead of paying per-token API prices. That works well for a single-user assistant with modest daily volume, but it constrains how the agent is invoked — you are routing through a subscription surface rather than a raw API endpoint. If you later add multiple users or heavy automation, compare the subscription route against metered API pricing honestly, because a busy agent can burn through tokens quickly.

Connecting your apps with Composio

Composio is the integration layer that turns "the assistant knows about your tools" into "the assistant can act on your tools." Instead of writing OAuth flows and API wrappers for each service, you register the apps you use and let the agent call them as tools.

  • Gmail for reading, searching and drafting messages.
  • Google Calendar for reading the day's schedule and proposing new events.
  • Slack for posting summaries and receiving commands.
  • Any additional service in the catalogue that exposes the actions you need.

Start with read-only scopes and add write permissions one at a time. An agent with send rights on day one is a liability; an agent that drafts and waits for approval is a colleague. The email reviewer workflow is the clearest example: it scans the inbox each morning, classifies messages, drafts replies for the ones that need them and leaves the send decision to you.

Giving the assistant a voice with ElevenLabs

The voice layer is optional but it changes how the assistant feels in daily use. ElevenLabs converts the agent's text output into natural speech, which is what makes a phone-style conversation possible rather than a chat window. Generate audio only for responses you actually want spoken — briefings, confirmations and short answers — because synthesising every log line is slow and expensive.

One expectation to set: voice replies are not instant. The pipeline has to transcribe or receive input, run the model, execute any tool calls and then synthesise audio, so a multi-second delay is normal. Editing that latency out of a demo video makes the assistant look faster than it is in production.

Common questions about always-on AI agents

Is the assistant really that fast? Response time depends on how many tool calls the model makes before answering. A calendar lookup is quick; a five-app workflow is not. Expect several seconds and design the interface around it.

Does it keep running when the computer is off? Yes — that is the entire reason for the cloud server. The agent lives on the host, not on your machine, so local shutdowns are irrelevant.

How much maintenance does it need? Less than a self-managed VPS but not zero: watch the logs, rotate keys, review what the agent did each morning and update the prompt when its behaviour drifts.

Can it be trusted with email? Only with guardrails. Draft-and-approve is the safe default; autonomous sending should be limited to low-risk categories you have explicitly defined.

The bottom line

A 24/7 JARVIS built on GPT 6 Astra is less about the model and more about the plumbing: a cloud server that never sleeps, a tool layer that reaches your apps, a prompt that defines boundaries and a voice that makes it usable. The build is reproducible without writing code, and the full code is shared so you can clone it rather than start from a blank page. Treat the first version as a draft-and-approve assistant, test the reboot path, and only then grant it write access to anything that matters. This breakdown is based on the walkthrough published by Damian Malliaros, whose channel covers the full build end to end; if you want the agent online today, start your server through Cloudways and apply the discount code before the first bill.

Questions & answers

Do I need coding skills to build a GPT 6 Astra assistant like this?

No. The build is assembled from hosted services: a cloud server, a prompt, Composio connections and an ElevenLabs voice. The full code is shared so you can clone it, but you still need to paste environment variables and follow the deployment steps carefully.

Is the assistant's response instant, or is the delay edited out of the demo?

It is not instant. Every reply may involve model reasoning, one or more tool calls and speech synthesis, which adds up to several seconds. Demos often trim that waiting time, so judge the assistant by its real end-to-end latency.

Can I run it on my existing GPT subscription instead of paying API prices?

The build keeps the agent on an existing GPT subscription rather than metered API usage, which keeps costs predictable for a single user. If you add users or heavy automation, compare that route against API pricing before scaling.

What should the agent be allowed to do on its own?

Start with read-only access to calendar and mail, and let the assistant draft replies rather than send them. Add write permissions one at a time once you have reviewed its output for a few days and trust its judgement.

Source: I Built My Own JARVIS with GPT 6 Astra (Runs 24/7) · Markdown version