taste-ai

taste is an open-source, self-hosted CLI utility that packs AI agent context — compressing long session histories and automatically extracting coding style from git commits — into a compact file ready for reuse across agents. The core workflow is three commands: navigate to your project, run `taste`, then open your agent with the generated context. The vendor describes the binary as framework-agnostic, meaning the output file can be passed to any AI coding agent that accepts a context or system-prompt file.

The differentiating feature is the automatic learning step. Rather than asking you to write a coding guidelines document by hand, taste reads your git history to infer your actual patterns — variable naming, error-handling structure, module organization — and folds those into the compressed context. This matters when you are switching between multiple agents on the same codebase and need each one to behave consistently without re-prompting style rules from scratch every session.

taste fits tightly into workflows where the project has accumulated enough git history to make pattern extraction meaningful. The tool installs via a single curl command, carries an MIT license, and has no paid tier or cloud dependency — the entire operation runs locally on the machine where you invoke it. Where it breaks down: a greenfield repository with few commits gives the extraction step nothing to learn from, leaving the compressed context thin and potentially less useful than a hand-written prompt. Teams in that position end up writing explicit style guides manually, at which point the auto-learn value proposition is gone.

The GitHub repo ships a `taste` binary, an `install.sh` script, and an `.agent-taste.json` configuration file in the repository root. No API is exposed and no external service calls are described in the source material — the tool processes local files only.