Pricklypear

Pricklypear is a multi-user Lisp operating system. One unix process is the whole OS surface — HTTP, REPL, eval, auth. Postgres is the filesystem: every function, type, route, and data row lives there as rows. The image is ephemeral and rebuilds from the store.

You don't deploy Nopales code. You save a row. Change a function and every route that calls it picks it up immediately — no compile step, no restart.

Nopales is the language. It's a small Scheme-family Lisp-1. It's homoiconic: the AST is the source, and function definitions persist in Postgres as JSONB. Editing the AST is editing the code.

Here's a handler function. The browser hits /hello and this runs.

(define hi-page
  (lambda () (render-html "<h1>Hello</h1>")))

You don't need to restart to see this page. There's no translation layer. Postgres just stores the direct AST of your sexp, serialized as JSONB:

{"lambda":[[],
  {"app":[{"var":"render-html"},
            {"lit":{"str":"<h1>Hello</h1>"}}]}]}

It has closures, lambda, let, quasiquote macros, pattern matching, and a frozen core. Dynamic typing, impure functional, effects-inferred. Two kinds of function: data builders you call from the REPL, and HTTP handlers you bind to a URL.

A handler is just (route "/hello" "hi-page") — where hi-page is a function you defined. The browser hits /hello; the REPL calls hi-page.

After you log in (or from scripts/repl.sh against a local image):

1. scripts/repl.sh
2. (define hi-page (lambda () (render-html "<h1>Hello</h1>")))
3. (route "/hello" "hi-page")
4. open /hello in the browser
5. edit at /edit/hi-page
6. optional: (propose-function "double" "lambda of int x -> (* x 2)")

Remote agents (Agora, pi, curl on a phone) do not need a custom REST surface. Fetch the public bootstrap doc, take a username/password from a human once, then POST Lisp to /api/eval.

Machine contract: /.well-known/agent.json · Prose: /agent.txt. No secrets in either document.

Has anybody noticed there's a third thing missing? I call it The Third Substrate if I'm trying to be fancy. Your harness writes code. Your agent manages tasks and processes. But there's nowhere to deploy your code. Now your poor OpenClaw is a TypeScript DevOps engineer

Pricklypear is the third thing. It's an living app and a memory. If you want you can write all your agent memory to it.

The agent calls it a habitat. I kind of think that name sucks, but oh well. Maybe its natural habitat is living inside its own Lisp. I think Cloudflare is trying to solve this problem with some insane, poorly designed abstraction. (like everything they touch)

I kind of think a habitat has to be a Lisp. I think you could build it in CL or Clojure. But there are big advantages to having an interpreted Lisp. I'd actually like to see a CL or Clojure habitat. My agents say you could build one in less than 10000 LoC. That's about a quarter of the size of Pricklypear. But then I built an interpreter, too.

Browse the source or download the book without an account. To run code in the image, log in — workspace home, Functions, Pages, Data, and the REPL stay private.