Developer Builds Scheme Compiler to WASM Using AI in 4 Days

Matthew Phillips built Puppy Scheme, a Scheme compiler that compiles to WebAssembly, in approximately four days with AI assistance. The project was inspired by seeing others quickly build near-production-level tools that would normally take months or years.
Key Features and Capabilities
The compiler includes:
- Support for 73% of R5RS and R7RS Scheme standards
- Support for WASI 2 and the Component Model
- Uses WASM GC (garbage collection)
- Pretty good dead-code elimination, creating small binaries
- Self-hosting capability - Puppy compiles its own source code to puppyc.wasm
- A wasmtime wrapper to create native binaries
- A website that runs on Puppy wasm in Cloudflare Workers
Performance Improvements
One night, Phillips told Claude to "grind on performance" and when he woke up, compilation time had improved from 3½ minutes down to 11 seconds.
Example Code
The project includes a component model example:
(define count 0)
(define (counter-view)
(html
(div (@ (class "counter"))
(button (@ (on "click" "on_decrement")) "-")
(span (@ (class "count")) ,(number->string count))
(button (@ (on "click" "on_increment")) "+"))))
(define (handle-event handler)
(cond
((equal? handler "on_decrement")
(if (> count 0)
(set! count (- count 1))))
((equal? handler "on_increment")
(set! count (+ count 1)))))
Current Status
The compiler is still alpha quality, and Phillips runs into bugs frequently, so it's probably not ready for anyone other than him to use. However, he notes that "the number of features is impressive for how little time I spent on it." The project demonstrates how AI can accelerate development of complex compiler projects.
Phillips worked on this as a side project, spending most of a weekend plus an hour or two on a couple of weekday nights. He's not sure where he'll take it from here, but the project is available at puppy-scheme.org.
📖 Read the full source: HN AI Agents
👀 See Also

Blindspot MCP: An External Brain for AI Coding Agents
Blindspot MCP is a tool that indexes full codebases using tree-sitter and SQLite to help AI coding agents understand symbols, dependencies, and relationships across files, preventing changes that break code outside their immediate context.

OpenClaw Benchmark Shows Qwen3.5:27B Outperforms Other Local LLMs for Agent Tasks
A benchmark of 7 local LLMs on 22 real agent tasks using OpenClaw found qwen3.5:27b-q4_K_M scored 59.4%, while the runner-up qwen3.5:35b scored only 23.2%. Most models couldn't find basic tools like email functions.

Agent Memory Protocol (AMP): Open Spec for Interoperable AI Agent Memory on Top of MCP
AMP defines a standard interface for persistent memory in MCP-compatible agents with six core verbs: encode, recall, forget, consolidate, pin, and stats. Includes compliance test suite and reference implementation.

IUM: MCP Symbol Indexer Cuts AI Agent Token Usage by 15.9x vs grep
IUM indexes codebases into an SQLite matrix of symbol events, exposing exact file:line coordinates, call graph tracing, and semantic search via MCP. Benchmarked against DataFusion (1,538 files) showing 15.9x fewer tokens than grep for equivalent queries.