pycomet/agent-atlas

Agent Atlas is a tool that scans your AI coding environment and draws an interactive mind map of everything in it. If you use tools like Claude Code, Cursor, or Codex CLI, you have probably accumulated dozens of skills, subagents, and servers over time. Agent Atlas reads your configuration files and session history, then visualizes your entire setup as a map where every skill and agent appears as a node. The nodes are grouped by purpose, sized by how often you actually use them, and colored grey if they never fire. The map answers practical questions that are hard to figure out otherwise. You can see whether your setup leans more toward engineering, writing, research, design, or ops. You can spot installed servers that load into every session but never get used, which means you are paying a hidden token cost for dead weight. You can also find overlaps where two skills do nearly the same job, or gaps where an entire capability area has zero coverage. Below the map, three diagnostic lists break down dead weight, overlaps, and gaps in plain text. Running it is a single command: npx agent-atlas-cli. It requires no account or configuration. It detects supported tools on your machine automatically, including Claude Code, Codex CLI, Cursor, ORGN CDE, and OpenCode. Each tool provides different levels of data. Claude Code and Codex CLI offer full usage data from session logs, while Cursor provides inventory only with no usage numbers. A --rough flag skips the optional API call entirely and uses keyword classification instead, so it works with no API key. Privacy is a central design choice. The tool is read-only and local-only, meaning it never modifies anything on your machine. The only data that leaves your computer is an optional classification call that sends just the names and descriptions of your skills and agents. Your session transcripts, code, and prompts never go anywhere. With the --rough flag, nothing leaves your machine at all. The project is open source under the MIT license, so you can inspect the code yourself. The pipeline behind it has four stages: a Scanner that inventories what is installed, a Usage Miner that counts what actually fired in your recent sessions, a Classifier that scores each item across five capability axes, and a Renderer that draws the map and diagnostics. Classification results are cached by content hash, so repeat runs are fast and cheap. You can also manually override any classification the system gets wrong.