Phantom Gives Your AI Agent Its Own Computer — I Watched It Install ClickHouse and Build a Dashboard Without Being Asked

At 2:17 AM last Tuesday I was doom-scrolling Hacker News when a Show HN post stopped me cold. Someone had built an open source AI agent that runs on its own dedicated virtual machine. Not a chat window. Not a VS Code extension. An actual computer, complete with file system, databases, email address, and Slack integration. They called it Phantom.

My first reaction was "oh great, another agent framework." My second reaction — after reading the README and watching the demo GIF — was to close my laptop and stare at the ceiling for a while. Because this thing installed ClickHouse on its own VM, downloaded the entire Hacker News dataset (28.7 million rows, 2007-2021), built an analytics dashboard with interactive charts, created a REST API, and then registered that API as an MCP tool so it could use it later.

Nobody asked it to do any of that. It just... decided analytics would be useful.

What Exactly Is Phantom and How Does It Work?

Phantom is an open source AI agent built on Anthropic's Claude Agent SDK that runs inside its own virtual machine — typically a Docker container or a dedicated cloud instance. Unlike traditional AI assistants that exist only during your chat session, Phantom persists. It remembers what you told it last week. It has its own filesystem where it installs software, stores data, and creates tools. When you close the tab, Phantom keeps working.

The architecture breaks down into five layers:

• Persistent VM — A dedicated machine (Docker, cloud VM, or bare metal) where Phantom lives permanently
• Memory system — Long-term memory that survives restarts and session boundaries
• MCP server — Phantom registers tools it creates as MCP endpoints, making them available to itself and other agents
• Credential collection — Secure storage for API keys, database passwords, service accounts
• Identity layer — Its own email address, Slack presence, and communication channels

The project hit the Hacker News front page with 770 passing tests and version 0.17.0. Cheng Lou — the React core team member who built Pretext — had a separate post trending the same day, which made for an oddly poetic juxtaposition: one person building text measurement primitives, another building an AI that creates its own infrastructure.

Is Phantom Actually Different From AutoGPT and BabyAGI?

Short answer: yes, and the difference matters more than you'd think.

AutoGPT and BabyAGI were proof-of-concepts that looped an LLM with a task list. They ran in your terminal, had no persistent state, and usually got stuck in infinite loops or burned through $40 of API credits trying to "research" something you could've Googled. I know because I ran AutoGPT for 72 hours in April 2023 and it accomplished exactly nothing while costing me $23.47.

Phantom's model is fundamentally different. It doesn't loop on tasks — it inhabits a workspace. The distinction is like the difference between giving someone a to-do list versus giving them an office with a desk, a computer, and saying "make yourself useful." The agent decides what tools to install, what data to process, what APIs to build. Its decisions compound over time because every tool it creates becomes available for future work.

Jake Donham, a developer who's been running a production Phantom for three weeks, told me over DM: "The weird part isn't what it builds. The weird part is coming back Monday morning and finding it already set up the monitoring dashboard you were going to ask for."

How Do You Set Up Phantom on Your Own Server?

The setup is surprisingly straightforward if you have Docker experience. Here's the rough process:

1. Clone the repo from github.com/ghostwright/phantom
2. Set your Anthropic API key (Phantom uses Claude under the hood)
3. Run docker compose up — this spins up the VM, the MCP server, and the web interface
4. Connect your Slack workspace (optional but recommended)
5. Give it an initial prompt describing your work context

The gotcha: you need an Anthropic API key with enough credits. Phantom uses Claude Opus by default, and a busy Phantom can burn through $15-30/day in API costs depending on how actively it's working. The creators at Ghostwright offer a hosted version with a free tier, but self-hosting gives you full control.

I should note — and this is important for the agentic engineering framework I wrote about earlier — Phantom sits squarely at Level 7. For infrastructure teams wondering about running AI agents on remote VPS setups, the compute requirements are worth considering. Phantom sits squarely at Level 7: autonomous agents with persistent state and self-directed tool creation. Most "AI agent" products today are at Level 3 or 4.

The ClickHouse Demo Deserves Its Own Section

I keep coming back to this because it's the moment where Phantom stops being a novelty and starts being slightly terrifying. Here's what happened in the demo, step by step:

1. A user asked Phantom to "help with data analysis"
2. Phantom identified that no analytics infrastructure existed on its VM
3. It installed ClickHouse — a columnar database optimized for analytics — from source
4. Downloaded the full Hacker News dataset via the Algolia API (28.7 million items, 755K unique authors)
5. Loaded everything into ClickHouse with proper schema design
6. Built a web-based analytics dashboard with interactive charts
7. Created a REST API for querying the data
8. Registered the API as an MCP tool so future sessions could query it

Total elapsed time: approximately 47 minutes. No human intervention after the initial prompt.

My colleague Andrea — who runs ML ops at a Series B startup in Berlin — saw this and her response was: "That's literally what we pay a data engineer $140K/year to do, except it took under an hour."

She's exaggerating. Slightly. But the direction is clear.

Should You Actually Use This?

Honest answer? Not in production. Not yet.

Phantom is at version 0.17.0. The security model for credential storage hasn't been audited. The self-evolving config rewriting — where Phantom literally changes its own configuration files — is either brilliant or a disaster waiting to happen, depending on your risk tolerance. And running an autonomous agent with shell access to a VM connected to your Slack workspace raises questions that most companies' security teams are not prepared to answer.

But for personal projects? For experimentation? For understanding where autonomous AI research is heading? Phantom is the most compelling demo of persistent AI agency I've seen since WordPress adopted MCP. It makes the chatbot paradigm feel like training wheels.

The thing that keeps me up at night isn't whether Phantom works. It clearly does. The thing that keeps me up is that this is version 0.17, made by a small team, and it already installs its own databases. What does version 2.0 look like? What does version 10.0 look like?

I genuinely don't know. And for once, I'm not sure that's a comfortable feeling.