agent-tack

Dotfiles for AI coding agents (formerly agent-harness). One install gives Claude Code, Codex, Copilot, Cursor, Gemini, OpenCode and Crush the same instructions and skills (Cursor needs its global rules pasted once), so they work like a disciplined senior engineer in every project. Git-level safety (commit rules, secrets, protected main) applies to every tool; agent-level enforcement runs in Claude Code and Codex, and the command guard in Cursor (what each tool gets).

What it does

• One workflow, scaled to the task: plan, test first, small commits, docs, review; light for a typo, thorough for a risky change.
• Rules that are enforced, not just suggested: git hooks (every tool) and the Claude Code and Codex hooks block AI attribution, committed secrets, force-pushes to main, dangerous commands and merges with red or unfinished CI.
• Opt-in per project: everywhere else the assistant works normally with only the safety net on.
• Yours to change: rules, skills, modes and toggles are plain files and commands. Why →

Quick start

Requirements: Linux, macOS or Windows with WSL2; git, bash, python3 or jq; at least one supported AI tool.

git clone https://github.com/nbfrodri/agent-tack.git ~/Projects/agent-tack
~/Projects/agent-tack/install.sh      # keep the checkout there: installed files link to it

cd ~/Projects/my-app
tack enable                           # turn the workflow on for this project (--scaffold: add the docs skeleton too)

Restart your AI tools, then work as usual: "Add Google login", "Fix issue #12", "What would you improve?", "Prepare a release".

Preview with ./install.sh --dry-run (the first install lists every hook it adds; --no-hooks adds none), check with tack doctor, update with git pull && ./install.sh, remove with ./uninstall.sh. Installation details → · Editors and WSL →

Modes

The mode decides how much process each task gets. auto is the default and picks a level per task.

| Mode | For | What the assistant does | | --- | --- | --- | | lean | Small tasks when tokens matter most | Minimal self-contained rules, terse replies; still branch, commit and test (measured: about lite cost, ~10% less on new projects) | | lite | Typos, config, one-line fixes | Branch, commit, a test when logic changes | | standard | A bounded feature or bug fix | Adds TDD, a short plan and the affected docs | | strict | Risky or multi-module work | Adds a saved plan you approve, handoffs, review and delegation | | unleash | Unattended work on a branch | Works without asking; the guard still blocks dangerous commands (risks) |

tack mode strict            # this project
tack mode lite --global     # your default everywhere
tack mode new spike --from lite   # your own mode, in ~/.config/agent-tack/modes/

Modes in detail →

Everyday commands

| Command | What it does | | --- | --- | | tack status | Is the workflow on, which mode, is the formatter trusted | | tack config | List and change feature toggles (delegation, fast check, limits…) | | tack doctor / --tools | Diagnose the installation / each installed AI tool | | tack log | What the hooks did recently (opt in with tack config activity-log true) | | tack trust | Allow automatic formatting and the fast check in this checkout | | tack disable | Turn the workflow off for this project |

Run tack help for everything. Usage →

What's included

• Skills: a core workflow (dev-workflow) plus process skills (debugging, testing, releases, issues, reviews, delegation) and stack skills (frontend, APIs, databases, auth, deployment…).
• Agents: planner, implementer, reviewers (code, security, performance, architecture, UI), test and docs writers.
• Hooks: commit conventions, secret scanning, a command guard (including green-only merges), auto-format, a fast check after edits, a check before the assistant stops and an opt-in activity log.
• Claude Code mods: a usage band with the active tack mode and your 5-hour and weekly limits, and a live pane of what the agent is doing.

Full list → · How it works →

How it compares

Other ways to shape an AI coding agent, and where agent-tack differs (as of October 2026):

| | A hand-written AGENTS.md / CLAUDE.md | Large plugin collections (e.g. ECC) | agent-tack | | --- | --- | --- | --- | | Approach | Instructions per project | A broad catalog: hundreds of skills, agents and commands you pick from and invoke | One opinionated workflow applied automatically, scaled per task by mode | | How you use it | The model reads it | You start workflows with commands (/plan, /code-review…) and install the parts you need | tack enable, then ask as usual; the mode decides the process | | Enforcement | None | Claude Code hooks, configured with environment variables | Git hooks (any tool, even manual commits) plus a command guard in Claude Code and Codex (Cursor: the command guard); toggles per project with tack config | | Tools | One file per tool | Claude Code first, adapters for many others | Claude Code, Codex, Copilot, Cursor, Gemini, OpenCode and Crush from one install (Cursor's global rules pasted once); rules in AGENTS.md, agent hooks in Claude Code and Codex, the command guard in Cursor | | Context cost | What you write | Grows with what you install; selective install recommended | Small core; documents and skills load on demand | | Footprint | None | Node.js runtime and packages | Bash and git, plus python3 or jq |

Pick a catalog when you want breadth and choose workflows yourself; pick agent-tack when you want the same disciplined process in every project and tool, enforced by git everywhere and by agent hooks where the tool supports them, with little setup. They are not meant to be installed together: both add session hooks and overlapping rules.

Results

Measured on real sessions with Claude Sonnet 5.5 in Claude Code, plain assistant versus tack. Other models and tools can give different numbers (method and limits):

| | Plain assistant | Lite / Auto | Strict | | --- | --- | --- | --- | | Bug fix on a branch with a fix: commit and a regression test | 0/2 | 2/2 | 2/2 | | Cost of a bug fix | 1× | ~2× | 3.5× | | Cost of a new project | 1× | ~2.2× | waits for your plan approval |

Learn more

| | | | --- | --- | | Usage | Modes, toggles, startup context, delegation, overrides | | Editors and AI tools | VS Code, Cursor, JetBrains, each AI tool, WSL | | Customization | Change rules, skills, modes, guard rules and supported tools | | How it works · Architecture | Installer, hooks, components and flows | | Components · Conventions | Every skill, agent and mod; commit, PR and release rules | | Results · Development | Benchmarks; tests, evals and contributing | | Sharing · AI usage | Using it on another machine; how AI builds this project |