oml
oml.shcoding agentbabashka

A minimal coding agent you change while it runs

oml is a coding agent in Clojure: its own agent loop in 1481 lines of babashka. Every step it takes is a named function, and you change any of them while it runs.

Get startedgithub.com/sm-th/oh-my-lisp
oml running in Toad: tool cards, a /eval that changes the system prompt, and the next answer in pirate voice
oml in Toad: tool calls as cards, then one /eval changes the system prompt for the next answer.
Demo video. Open it directly if it does not play here.

What it is

  • Its own agent loop, small enough to read. Stream a Chat Completions call, run the tool calls, append the results, repeat. Built-in tools: read, write, edit, bash.
  • Everything is a function. Each step of the loop, each tool and each slash command is a named function the core calls through its var. Redefine it with defn, wrap it with advise!, set a setting with setq: from an init file, from /eval, or from an nREPL connected to the running agent. The next call uses it.
  • ACP is the UI. oml speaks the Agent Client Protocol over stdio, so Toad, Zed and Emacs agent-shell are its interface.
  • Any OpenAI-compatible endpoint. A base URL, a key and a model id; responses stream over SSE.

Quick start

With nix (flakes enabled):

git clone https://github.com/sm-th/oh-my-lisp oml
cd oml
cp .env.example .env      # set OPENAI_BASE_URL, OPENAI_API_KEY, OPENAI_MODEL
nix develop               # or: direnv allow
oml-toad ~/some/project   # Toad as the UI, oml as the agent

Then, in the session, try /tools, /doc oml.agent/run, or change the agent while it runs:

/eval (oml.custom/advise! #'oml.agent/system-prompt :pirate
        (fn [system-prompt] (str (system-prompt) "\n\nAnswer like a pirate.")))

Read on

  • Getting started: install, endpoint, Toad, Zed, Emacs, print mode, tests.
  • Customising: defn, advise!, setq; tools and commands as functions; init files; nREPL.
  • Primitives: the session, its transcript, say, complete, call-tool, request!.
  • Recipes: project context, usage, search tools, /model, permissions, a different loop.
  • Architecture: the namespaces, one turn from ACP to the model and back, limits.
  • Chapters: what is done and what is planned.