Zerostack v1.5 – Lightweight Unix-inspired coding agent

Zerostack released v1.5 of its lightweight Unix-inspired coding agent, achieving feature parity with Claude Code on key agent capabilities. The update introduces a redesigned TUI, multi-modal input support, advisors for cost-efficient model routing, and configurable chain-of-prompts, along with various improvements and bug fixes.

With v1.5, Zerostack has achieved feature parity with Claude Code on the most important agent capabilities, while shipping a new and improved TUI. Additions: - Fully re-designed TUI w/ new input bar, new status bar & new welcome message; scroll down for screenshot - Added Multi-modal inputs with support for images, audio and PDF compile gated - Added Advisors, allowing to use cheaper models that can call "expert" LLMs when the task is too difficult - Added /review , which sends a specialized message to the agent based on your current Git context - Added configurable chain-of-prompts, allowing to go from brainstorm to plan to code and ending with /review - Added opt-in pre-configured MCPs for Context7 and Grep.app - Added --load-prompt and --pure-stdout , allowing for better scripts and tools built on top of zerostack - Added support for signaling agent's activity via UNIX sockets, allowing for better scripts and tools built on top of zerostack - Added installations via scripts, Nix and Homebrew - Added special permissions to allow plan prompt to write PLAN .md files Improvements: - Added ability to select directories using @ - Improved color rendering - Better tool use TUI rendering - Improved compaction - Improved doom loop detection - Improved caching on list of models - Added more configuration options - Enabled caching on Antrophic models thanks @xavierforge https://github.com/xavierforge - Several optimizations for local LLMs, with more coming during the v1.5.x cycle thanks hal-eisen https://github.com/hal-eisen - Fixed several bugs in git worktree module