Conversational CTF agentConcept
A capture-the-flag game where players talk to an AI hacker partner that runs real tools inside an isolated lab network. Concept and architecture stage.
Most CTFs are solo command-line work. This one puts a conversational agent between the player and the tools: the player supplies strategy and critical thinking, the agent translates intent into tool calls and narrates results.
Architecture
A web chat UI talks to a game API that tracks each player's progress. The API hands prompts to a small local model on the Jetson with a whitelisted set of security tools exposed through MCP. The tools only reach an isolated network of intentionally vulnerable machines, with a flag buried on the last one.
Why a small model is enough
The agent does not need to be brilliant. It needs to reliably turn "scan the network for open ports" into the right tool call and explain what came back. That is squarely what quantized 2B to 3B models with good tool-use training can do at the edge.