Autolith: a coding agent that lives inside a self-modifying Lisp REPL
Open-source and ISC-licensed: a live SBCL image the model rewrites on the fly, reads repos bigger than its context window, and runs on any endpoint you bring.
Copy markdown
The prompt box is the REPL
Autolith runs inside a live SBCL 2.6.6 image where prose and executable Lisp share one input, and the agent redefines functions and classes as it works — every change written to an append-only mutation log it can roll back. One curl line installs it on Linux, macOS or the BSDs.
Bring your own model — default is GPT-5.6 Sol
It talks to any OpenAI-compatible endpoint — Anthropic's API, Codex/ChatGPT, Grok, Fireworks, OpenCode — so you pay your existing provider and swap models with a flag. The default is gpt-5.6-sol, the tier OpenAI just cut to $4/$20 per million tokens.
It reads codebases bigger than the context window
The headline trick is “recursive inference over corpora larger than the model window” via content-addressed object interning: point it at a repo that won't fit in context and it chunks and recurses for you, no bespoke RAG plumbing. Conversations, memories and agendas persist as readable append-only files with a crash-recovery vault.
The catch: it runs the model's code as you
Autolith executes model-generated Lisp with your own user privileges; its docs warn plainly that “sandboxing is no substitute for human oversight.” Treat it as a power tool on a scratch repo, not an unattended autopilot. It's at v0.35 and on the Hacker News front page today.