Nodecode — an open-source Lisp agent that rewrites itself

The open-source Nodecode agent in Common Lisp rewrites itself — the model has a single tool, eval, and on ARC-AGI-3 it uses 54% fewer tokens than Claude Code.

Compiled by Sergey KostenchukPublished 2026-10-26Updated 2026-10-26

2026-10-26 Coding Anthropic

🛠 Nodecode — an open-source Lisp agent that rewrites itself

The model has one tool — eval, which executes Common Lisp directly in a running SBCL process. The agent saves skills and MCP connectors as .lisp files in git and loads them on every startup, while the terminal, browser, phone, and Discord/Telegram/Slack bots work through a single gateway process.

🌍 This is an alternative to the harness architecture of Claude Code and Codex CLI, where memory is scattered across markdown files. On ARC-AGI-3, Nodecode on Claude Opus 5.5 scored 100 points, using 54% fewer tokens than Claude Code, while GPT-6.1 Sol won all 25 games.

👤 It installs with a single line on Linux/macOS/Windows and is licensed under MIT. The agent writes missing tools itself and remembers them after a restart. There is no sandbox: eval executes the shell with user privileges, so a container or VM is required for autonomous operation.

Source 1: https://github.com/nodecode-ai/nodecode Source 2: https://news.ycombinator.com/item?id=50238751

Sources