The best coding agent harnesses, according to hackers
11 coding agent harnesses recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.
All coding agent harnesses
- Pi coding agentA minimal coding agent harness that works with many models, uses fewer tokens and is easy to extend, though basics like subagents are missing and it has no sandboxing.892recommend
- omp (oh-my-pi)Free terminal coding agent that lets you switch models freely and has good defaults. Critics call it bloated and say subagents burn usage quickly.80recommend
- ForgeCodeFree coding agent harness that people credit with reliable tool calling for small local models and strong Terminal Bench results; critics question its benchmark claims.39recommend
- ReasonixRecommended for cutting costs of long DeepSeek sessions via cache-friendly design; the most common caveat is a rough terminal interface.26recommend
- DSHFree coding-agent harness people use as a TUI and chat client, praised for extensibility, configurability and power on the same underlying model. One critic calls it incomplete.11recommend
- smolA minimal open source coding agent that is small, light on CPU and RAM, and easy to read and adapt. Critics note it lacks subagents and a system prompt, and its size hurts caching.9recommend
- Zot coding agentFast single-binary terminal coding agent with modest memory use and no gateway overhead, used daily by several people. Critics say it spoofs Claude Code requests to the Anthropic API.9recommend
- Pu.shRecommended as a small shell-based coding agent that fits in a few hundred lines. Critics say the code is minified to hit the line count and is hard to read.8recommend
- little-coderRecommended as a minimal coding agent built on pi, with defaults and fixes aimed at running small local models.7recommend
- jcodeFree open-source Rust coding agent harness that people praise for low memory use, efficiency and browser integration. One critic calls its design peculiar.6recommend
- BenziCoding agent that compiles a codebase into deterministic lookups, so it reads far less source code and still scores well on SWE-bench.5recommend