Pu.sh: Minimal shell-based coding agent framework in 400 lines

What it does

Pu.sh is a shell-based coding agent framework written in approximately 400 lines of code. It accepts tasks and generates code or solutions by calling an LLM API. The tool handles tool-use loops, reasoning steps, and JSON parsing entirely within POSIX shell, curl, and awk—no external packages required.

Who it is for

Developers who need a portable coding agent that runs on minimal systems. The target is those working in constrained environments: embedded systems, remote servers, or CI/CD pipelines where adding dependencies is difficult or undesirable. It appeals to builders who prioritize simplicity and portability over feature richness.

Pricing

The site does not show prices. Pu.sh itself is open-source and free to download and run. Users must supply their own OpenAI API key, so costs depend on LLM usage.

How it stands out

The core constraint is deliberate minimalism: zero package dependencies, under 50KB total size, POSIX shell only. Most coding agents rely on Python, Node, or Docker—adding layers of dependency management and setup friction. Pu.sh eliminates that entirely. It works anywhere a shell, curl, and awk are available. The implementation includes genuinely unconventional techniques, such as JSON parsing and tool-use reasoning loops implemented in awk rather than a traditional language.

What a founder should check

First, verify the actual user experience in interactive mode. The launch text notes the first cut was "unusable interactively"—confirm whether that limitation persists and whether it matters for the intended use cases.

Second, assess switching costs from existing agents. Projects already integrated with Python-based tools like Anthropic's or OpenAI's official libraries may find the shell overhead offsets the portability gain. Measure the friction of rewriting agent logic in awk versus the benefit of dropping dependencies.

Third, understand the LLM cost structure and token efficiency. Shell-based agents may struggle with complex reasoning tasks, driving higher API costs. Verify whether the framework reliably handles stateful interactions or whether it wastes tokens on formatting and parsing overhead.

Thinking of building something like this?

Every launch here is a competitor to somebody's idea. If yours is close, check it against the market before you build: the Full Check names the rivals, the prices and the gaps.