Agent Smith
Point it at your code. It reads your project, then sets up your AI coding assistant to work your way — with the right commands, the right helpers, and guardrails so it can't make a mess.
Explain it like I'm five
Imagine you get a brand-new robot helper. Out of the box, the robot is smart but it doesn't know your house — where the kitchen is, which cup is yours, that you never wear shoes inside.
Agent Smith is the person who shows the robot around your house. It walks through your project, notices how things are done ("ah, this is a Java kitchen, the tests live here, you tidy up with this tool"), and then hands the robot a little instruction card so it helps you the way you already work — instead of guessing and doing it wrong.
It also puts up a few safety gates ("don't touch the stove," "always wash the cup after") so the robot can help on its own without breaking anything. You're still in charge — the robot asks before doing the big, scary stuff.
That robot is Claude Code (Anthropic's AI coding assistant). Agent Smith is the setup crew that makes Claude Code instantly useful in your repository.
The problem it solves
A general AI assistant doesn't know your project. So it:
- runs the wrong commands (Python's
pytest on a Java project — a real bug this tool was built to kill),
- ignores your conventions (your folder layout, your auth rules, your logging style),
- has no memory of how the code is wired, and
- has no guardrails, so letting it run freely is risky.
You could hand-write all that setup yourself for every repo. Agent Smith does it for you, automatically, by reading your actual project — and keeps it honest: if it can't tell what something is, it says so instead of guessing.
What you get after running it
One command turns a plain repository into a Claude-Code-ready workspace with:
| Thing it installs | What it's for |
|---|
Slash commands (/as-backend, /as-test, /as-ship, …) | One-word shortcuts for everyday jobs, pre-loaded with your stack |
| Skills (review, test, docs writers — plus fable-mode) | Detailed playbooks the assistant follows for specific tasks |
| MCP servers (gitnexus, git-memory, serena, …) | Give the assistant memory: code structure, git history, symbol search |
| Hooks (session start, pre-tool, stop) | Automatic checks that run around the assistant's actions |
| Sentrux quality gate | A guardrail that blocks changes which make the architecture worse |
A managed CLAUDE.md section | A living cheat-sheet of every command and skill, refreshed on each run |
All of it is tailored to the stack Agent Smith detected — not a generic template.
Quick start
You need Node 20+ and the claude CLI on your PATH (for the smartest setup; it still works without it, just less customized).
# from the root of the project you want to set up
npx @gunesbizim/agent-smith init
That's it. Restart Claude Code and try a command like /as-backend "add a health endpoint".
Want to look before you leap?
npx @gunesbizim/agent-smith analyze # just print what it detects, change nothing
How it works (the four steps)
┌───────────┐ ┌───────────┐ ┌──────────┐ ┌──────────┐
your │ 1. DETECT │ ──▶ │ 2. ADAPT │ ──▶ │3. INSTALL│ ──▶ │4. OPERATE│
repo │ the stack │ │ skills & │ │ MCPs, │ │ commands,│
│ │ │ docs to it│ │ hooks, │ │ guarded │
└───────────┘ └───────────┘ │ gate │ │ pipeline │
└──────────┘ └──────────┘
1. Detect — read the project, never guess
Agent Smith gathers evidence the project declares about itself — build manifests and CI files (pom.xml, build.gradle, package.json, go.mod, Cargo.toml, pyproject.toml, .github/workflows/…, Makefile, …) across every module. It then synthesizes a StackProfile: the language, framework, ORM, database, and the real commands to test/lint/format/migrate.
- If the
claude CLI is present, an LLM pass reads the evidence and classifies the stack — so it covers essentially any language without a hardcoded list.
- If not, a deterministic fallback handles the common ecosystems (Java/Maven+Gradle, Node, Go, Rust, Python).
- If something can't be determined, it is reported as
none — never filled in with a borrowed default. (This is the honesty rule that stops a Java project from being told to run ruff/pytest.)
