Developer Nahim Nasser released Pu.sh on April 30, 2026, a fully functional coding agent built in approximately 400 lines of shell script. The project uses only system primitives—sh, curl, and awk—with no additional dependencies required.
Developer Built Portable Agent With Self-Imposed Constraints
Nasser originally experimented with pi-autoresearch, giving it the task of building the most portable coding agent possible. The first iteration produced 6 KB of shell script that worked for one-shot tasks but was unusable interactively. Nasser then expanded the project with strict self-imposed rules: no new dependencies and under 500 lines of code.
The constraints led to unconventional implementation choices, including JSON parsing and OpenAI Responses API tool loop handling entirely in awk, with reasoning items carried across turns.
Agent Includes Seven Tools and Multiple Operating Modes
Despite the minimal codebase, Pu.sh includes substantial functionality:
- Support for both Anthropic and OpenAI APIs
- Seven tools: bash, read, write, edit, grep, find, ls
- REPL interactive mode
- Auto-compaction for conversation management
- Checkpoint and resume capabilities
- Pipe mode for command-line integration
- 90 no-API tests for reliability
Features explicitly not included are TUI, streaming, images, OAuth, and Windows support.
Architecture Inspired by Pi.dev With Exact-Text Edit Model
Nasser credits Pi as heavy inspiration for the project, adopting the same seven-tool surface and exact-text edit model. The agent loop itself is minimal, with most complexity coming from developer experience features and hardening.
Nasser acknowledged that AI models wrote much of the awk code, stating he cannot read most of it himself and that such a project would not have been possible for him a year ago.
Key Takeaways
- Nahim Nasser built Pu.sh, a complete coding agent in approximately 400 lines of shell script requiring only sh, curl, and awk
- The agent supports both Anthropic and OpenAI APIs with seven tools, REPL mode, auto-compaction, checkpoint/resume, and 90 no-API tests
- Implementation uses unconventional approaches including JSON parsing and OpenAI Responses tool loop handling entirely in awk
- The project was inspired by Pi.dev and demonstrates extreme portability with no npm, pip, or Docker dependencies
- Nasser credits AI models for writing much of the awk code and states the project would not have been possible a year ago