Phantom is an autonomous AI agent that operates on a dedicated virtual machine rather than users' local computers, eliminating the need to repeatedly explain context across sessions. Built by the Ghostwright organization on Anthropic's Claude Agent SDK, the open-source project has garnered 789 GitHub stars since its March 26, 2026 launch and currently passes 822+ automated tests in version 0.18.2.
The agent's core innovation centers on persistent operation: each Phantom instance runs continuously on a $7-20/month VM, maintaining memory and state between interactions through a three-tiered vector memory system powered by Qdrant.
Three-Tier Memory System Enables Cross-Session Recall
Phantom's memory architecture stores conversations, task outcomes, and learned preferences in vector format using Qdrant with Ollama embeddings. This persistent memory eliminates the common frustration of re-explaining requirements to AI assistants after each restart, allowing the agent to reference previous work and decisions automatically.
The system combines short-term context (current session), medium-term working memory (recent tasks), and long-term knowledge (accumulated expertise) to maintain continuity across days or weeks of operation.
Self-Evolution Engine Improves Performance Over Time
After each work session, Phantom evaluates its own performance and proposes configuration improvements. These suggested changes pass through multi-model LLM judges with "minority veto" safety gates before implementation, creating a feedback loop that refines the agent's capabilities without human intervention.
The system can also create and register its own MCP (Model Context Protocol) tools at runtime, extending its capabilities as new needs emerge during operation.
Production Demonstrations Show Real-World Capabilities
Phantom has demonstrated enterprise-grade capabilities across multiple domains. In one case, it built a complete analytics platform using ClickHouse that loaded 28.7 million Hacker News records with interactive dashboards. In another instance, the agent self-extended with Discord support when the original configuration lacked that integration.
Additional production use cases include infrastructure monitoring with Vigil (processing 890,450 metric rows), automated standup reports for engineering teams, infrastructure provisioning, data pipeline construction, and competitive intelligence monitoring.
Email Identity and MCP Server Integration
Each Phantom instance receives its own email address and provides public URLs for dashboards and APIs, functioning as a genuine team member rather than a transient tool. The agent operates as a full MCP server, connecting with Claude Code and other AI agents through the Model Context Protocol's streamable HTTP protocol.
Credentials are secured using AES-256-GCM encryption with magic-link authentication, addressing security concerns about autonomous agents handling sensitive data.
Key Takeaways
- Phantom runs on dedicated VMs ($7-20/month) with three-tiered vector memory using Qdrant, enabling persistent recall across sessions without re-explanation
- The agent's self-evolution engine evaluates performance after each session and proposes configuration improvements validated by multi-model LLM judges
- Production deployments include analytics platforms processing 28.7M records, infrastructure monitoring with 890K+ metrics, and automated data pipelines
- Each Phantom instance has its own email address and functions as a full MCP server, integrating with Claude Code and other agents
- Released under Apache 2.0 license with 789 GitHub stars, available for self-hosting via Docker Compose or through Ghostwright's free managed service