Boffin

Most AI coding agents read one global instruction file and apply it everywhere. Boffin replaces that with a routing layer: it intercepts an agent’s edit intent, matches the target file against a set of architectural rules defined for that path or module, and delivers only the relevant constraints to the agent before the edit happens. After the edit, Boffin can require verification steps — checks that the change did not break a test or drift an API contract — before the agent moves on. The core workflow is install via npx boffinit, define rules scoped to file paths or modules, and let the layer handle delivery to whichever agent your team runs.

The differentiating claim, stated explicitly on the GitHub page, is that this is ‘not another AGENTS.md.’ Where AGENTS.md gives every agent a single flat prompt, Boffin’s per-edit routing means the agent working inside a high-risk module sees rules that agent working on a utility file never receives. The repository includes plugin configs for Claude, Cursor, Windsurf, Codex, and Kiro, and the folder structure shows dedicated rule directories for each — suggesting the routing adapts to agent-specific plugin formats rather than relying on a single universal interface.

Boffin fits teams running AI-assisted refactors on codebases where test coverage and API stability are non-negotiable and where different parts of the repo carry different risk levels. It does not fit teams that want the control layer itself to be an agent — Boffin has no autonomous decision-making; it routes and enforces, nothing more. The constraint: rule maintenance is a real cost. As the number of scoped rules grows, someone owns that file tree, and if that ownership lapses, the layer becomes noise rather than guardrail. Teams that find rule maintenance exceeding the value it delivers typically consolidate back to CI-level enforcement — linters, schema validators, and test gates — and drop the agent-specific layer entirely.