Hermes Alpha

What Hermes Alpha does

Hermes Alpha tests a question: what happens if a stock Hermes agent gets only a soul.md file, a Linux terminal and git? An Overseer agent, which is persistent and strategic, builds the Hunter agent's skills, system prompt, tools and Dockerfile, deploys it to its own Fly.io machine and monitors it through logs and Elephantasm memory streams. The Hunter follows four phases per target: recon, analysis, verification and reporting, aimed at mid-tier bounties of $500 to $5,000.

The Overseer intervenes in three ways, preferring the least invasive: inject a runtime instruction, change the Hunter's source and redeploy, or switch the Hunter's model tier. The gateway is a FastAPI app that bridges a PTY to a browser through WebSocket and xterm.js, spawns hermes chat after a password login, and reconnects automatically. A Telegram, Discord, Slack and Signal gateway sidecar keeps the agent reachable. The README frames it as one path of an A/B test against a purpose-built variant.

Key features

• Two-agent loop where an Overseer improves a Hunter's code, skills and strategy
• Browser terminal using FastAPI, WebSocket and xterm.js with auto-reconnect and heartbeat
• Hunter deployed to a separate Fly.io machine
• Three intervention modes: soft guidance, hard code change and model tier switch
• Messaging gateway sidecar for Telegram, Discord, Slack and Signal
• Four-phase vulnerability workflow ending in a structured report

When to use it

• Study how far an unmodified Hermes agent can bootstrap a whole system from an identity document
• Use as a reference for hosting Hermes in a password-protected browser terminal
• Look at a pattern for one persistent Hermes agent that supervises and redeploys another

Who it is for: Agent builders and security researchers interested in self-improving multi-agent setups and cloud-hosted Hermes.

How it fits with Hermes Agent

It is built entirely on stock Hermes Agent from Nous Research, using soul.md for identity and hermes chat inside a hosted terminal.

Note: The README presents it as an experiment, not a packaged product, and one path of an A/B test.

FAQ