# claude-repo.com > Free, open-source tools for people who build with Claude Code and other AI > coding agents (Antigravity, Codex, Gemini CLI, Copilot). Each runs on your > own machine with no service of ours and no account with us. never-again and > never-collide are a script and a few files in your repository (bash, Python > 3.7+ and git); gsc-connector is a local MCP server (Node 20+). MIT. Built by > Malay Sherasia; not affiliated with Anthropic. The hub page at > https://www.claude-repo.com/ lists the tools; each has its own path below. - [never-again](https://www.claude-repo.com/never-again): mistakes that can't repeat. One agent stops repeating a bug you already fixed. - [never-collide](https://www.claude-repo.com/never-collide): agents that can't step on each other. Several agents share one repo without overwriting each other's work. - [gsc-connector](https://www.claude-repo.com/google-search-console-connector): Google Search Console inside Claude Code. Sign in with Google once, ask about clicks, queries and rankings, and export every report as CSV. Privacy policy: https://www.claude-repo.com/google-search-console-connector/privacy # never-again > A Claude Code skill that turns every bug you fix into a guardrail. It asks one > question per mistake — can this be enforced instead of remembered? — and > writes a hook if the answer is yes, or one capped line in LESSONS.md if it is > no. MIT licensed, local, no account, no telemetry. It talks to GitHub only > about releases: na upgrade when you run it, and a once-a-day check that > one setting turns off. never-again is installed into a repository, not run as a service. The whole tool is a skill definition, a shell hook template and a Python script that counts lines. No model is involved in its statistics. ## Who it is for - Non-technical builders who vibe-code or prototype with Claude Code and do not read the code line by line: when the AI fixes something, say "never again" and the fix becomes a rule it cannot ignore. Nothing to configure. - Developers: the same lesson becomes a hook that denies the wrong action before it happens, at the cost of one line, and the rules file stays short enough for the model to follow. Hooks also run from git, so any tool or terminal meets them. - Either way the saving is the same: time re-explaining, tokens re-reading, and the bug that would otherwise return next week. ## How it works - A bug is fixed, then the user says "never again" or Claude notices the correction. - Triage picks the first of four rungs that fits: an existing gate (turn it on, file nothing), hook-shaped (write a hook), judgement (one line in LESSONS.md), or not worth filing at all. - A mistake is hook-shaped if the check can be stated as one sentence beginning "before" or "after" — for example, "before committing, fail if the build was never booted in a browser". - Every new hook starts in warn mode, emitting permissionDecision "ask". What the person chose is recorded automatically (declined or proceeded); a declined fire counts as correct. After five correct in a row a person runs `na promote`, subject to the repo's "promotion" setting (pull request by default). Nothing self-promotes. - One registered entry, the dispatcher, runs every hook from an index and skips hooks whose watched file types did not change. The same hooks run from git's pre-commit, so commits from any tool or terminal are covered. - "X must pass before commit" (tests, a build, a browser boot) is one setting per lesson: the hook runs the command itself when watched files changed and fires only if it fails. - `na import` finds a repo's existing notes files (CLAUDE.md rules, NOTES.md, docs/lessons.md, editor rules) and the skill turns them into lessons without editing the sources. ## Key facts - Cost model: a rule line in an instructions file is paid for on every turn; a hook costs one line until it fires, because it lives outside the context window. - Enforcement: permissionDecision "deny" runs before every permission mode, including --dangerously-skip-permissions. - Density: LESSONS.md is capped at 40 rules per file by default, ordered most-fired-first by `na sort`, and scoped per package. The number that matters is rules loaded per turn, which `na` reports as "loaded here" and warns above 60. - Evidence for the cap: IFScale (Distyl AI, 2025, arXiv:2507.11538) measured 20 models from 10 to 500 concurrent instructions; the best reached only 68% at the top end, and earlier instructions were followed more reliably than later ones. - Privacy: zero accounts, zero telemetry. Two requests to GitHub, both about releases (na upgrade on request; a once-a-day newest-release check after a commit, off with "updates": "off"). fires.log is local and gitignored; LESSONS.md, the hooks and their modes are committed as team knowledge. A fresh clone runs the installer once to wire git's side. - Tested on Ubuntu, macOS and Windows on every pull request (168 assertions). - Latest release: 1.10.0 (2026-10-16). Release notes: https://github.com/malaysherasia-ai/claude-never-again/releases - Version 1.1.0 (2026-09-17): the same hooks run under Codex, Gemini CLI, GitHub Copilot and Antigravity (install.sh --agent NAME). - Version 1.0.1 (2026-09-16): file formats are frozen; changes to them bump the major version. ## Requirements - git, bash, and Python 3.7 or newer. - Works on macOS, Linux, WSL and Windows. On Windows, Git for Windows supplies bash.exe and PowerShell can call it directly; `command -v python3` there resolves to the Microsoft Store alias stub, which is why the interpreter resolver requires a candidate that actually executes a statement and exits 0. This was fixed in v0.1.1. Every Python the hooks spawn writes UTF-8, so an emoji in a commit message does not blind a hook (fixed in v0.4.0). ## Pages - [Home](https://www.claude-repo.com/never-again): what it is, how it works, install steps, and a comparison with putting the rule in CLAUDE.md. - [Docs](https://www.claude-repo.com/never-again/docs): the full README — triage, warn and block modes, the na CLI, monorepo scoping, and why the rule file is capped. - [Proof](https://www.claude-repo.com/never-again/proof): real na output and hook fire counts from the two repositories running never-again, pasted unedited and updated by hand. ## Source - [GitHub](https://github.com/malaysherasia-ai/claude-never-again) - [Discussions](https://github.com/malaysherasia-ai/claude-never-again/discussions) - [MIT licence](https://github.com/malaysherasia-ai/claude-never-again/blob/main/LICENSE) # never-collide > The second member of the never-* family, on the same site. never-collide > coordinates AI coding tools that share one Git repository: Claude Code, > Antigravity, Codex, Gemini CLI and Copilot working in parallel without > overwriting each other. Each tool claims the paths it is about to edit on an > append-only ledger (ledger.jsonl on the orphan branch agents/ledger), is > refused when another tool holds them, and records done and tested with the > time. A git fast-forward push is the lock. Enforcement is an edit hook in > Claude Code, a git pre-commit hook for every tool and terminal, and a > pull-request check (GitHub Action); every hook starts in warn mode and > `ncl enforce deny` promotes it. Claims expire (default four hours). There > is no service and no account; nothing is sent on its own (the installer > asks once per machine whether to send an anonymous count, Enter skips). > MIT licensed. Version 0.6.1 (2026-09-27). Commands: whoami, status, log, > claim, start, note, done, tested, release, handoff, check, enforce, > verify-pr, doctor, report, feedback, upgrade, uninstall. The edit hook is > wired and tested for Claude Code; registrations for Antigravity, Codex, > Gemini CLI and Copilot follow their documentation and are not yet > field-confirmed. - [Home](https://www.claude-repo.com/never-collide): what it is, an animated walkthrough of a claim, a refusal, a handoff and a release, a plain-language explanation, install steps, and a comparison with branches, CODEOWNERS and worktree isolation. - [GitHub](https://github.com/malaysherasia-ai/never-collide) - [The ritual](https://github.com/malaysherasia-ai/never-collide/blob/main/AGENTS.md)