The first coding agent discipline system

Your agent doesn't get to finish until the tests pass.

Docs

Safeword is the first coding agent discipline system for Claude Code, Codex, Cursor, and OpenCode.

It runs alongside your existing stack — never modifies your application code or runtime dependencies.

The Problem

Your agent ships fast. But it skips tests, ignores lint, and pushes whatever “looks right.”

Your senior engineers stop trusting it. Your vibe-coders can’t tell working code from confident garbage.

AI doesn’t decide when it’s done. You do.

The Fix

Automatic Enforcement — Every edit gets linted. Every phase exit needs an independent review. Your agent can’t skip it.

Strict Quality — Code, architecture, docs, dependencies. Nothing slips through.

Structured Discipline — BDD and TDD enforced. Debugging structured. Refactoring gated.

Human-friendly — Guardrails only fire during AI sessions. Your workflow stays the same. Use the linting in CI or precommit hooks if you want — your choice.

Install

bunx safeword@latest install

Works with Claude Code, Codex, Cursor, and OpenCode. Requires Node (22.22.3+, 24.16+, or 26.3+) and Bun. Auto-detects TypeScript, Python (beta), Go (beta), Rust (beta), SQL/dbt (beta).

What Happens

Safeword adds project configuration and linting setup to your repo, and delivers its skills and hooks through each agent’s native plugin — it never modifies your application code. Commit the project configuration — your team gets discipline the moment they pull.

Under the hood, your agent now moves through five phases — clarify, classify, build, verify, done — with gates that stop it skipping straight to “done.” See The Workflow.

Stay Current

bunx safeword@latest install