doorman11991/smallcode

SmallCode is a terminal coding assistant built specifically for small language models, the kind in the 8 to 35 billion parameter range that can run on a regular laptop or desktop instead of a cloud server. The README contrasts it with tools like OpenCode, which assume you are using a frontier model such as Claude or GPT-5 with a very large context window and reliable output. SmallCode tries to squeeze useful coding work out of weaker, local models by changing the design around their limits. You install it from npm with a single command, or grab a prebuilt tarball for Windows, macOS, or Linux that bundles Node.js so you do not need to compile anything. It then talks to a local model server such as LM Studio or Ollama, configured through a .env file. There is an optional escape hatch: if the local model fails after retries, SmallCode can hand off the task to a cloud model like Claude or DeepSeek, but only if you provide an API key. The README lists several design choices that target small-model weaknesses. A context budget engine caps tool output and summarizes old history so the model never overflows its window. A two-stage tool router shows the model only the tools it needs, which matters when the context is just 8 or 16 thousand tokens. A forgiving parser accepts tool calls in JSON, YAML, XML, or plain text, since small models often produce messy output. Editing uses search-and-replace patches rather than full file rewrites, because small models tend to truncate or hallucinate when reproducing whole files. There is also a planning layer that breaks complex tasks into a TODO file the model rereads each turn, plus detection for repetition loops and stuck patch attempts. The author claims an 87 percent benchmark score using a 4 billion active parameter model.