Works with any tool — Claude Code · Cursor · Codex · Windsurf / Antigravity · GitHub Copilot · Ollama Works with any stack — web · mobile · backend · scripts · data
Your AI assistant forgets your project every session, re-solves the same bugs, and sometimes ships code that quietly breaks security or UX. You pay for all of it in tokens and rework.
SmartStack is a small set of files you drop into your repo. It gives any AI coding assistant:
- 🧭 Persistent context — it always knows your project, so you stop re-explaining it. (~30–40% fewer tokens.)
- 🗂️ Indexed memory — it remembers past sessions and past bugs, looked up by index, not rescanned.
- 🛡️ Built-in standards — security, performance, and UX rules enforced before code ships.
- ⚡ Token discipline — read summaries, not the whole repo; keep tasks small; reason deeply only when needed.
It's just Markdown, JSON, and shell scripts. No install, no daemon, no network calls, nothing to trust.
SmartStack/
├── AGENTS.md # the portable "brain" every tool reads
├── adapters/ # one thin file per tool → points it at the brain
│ ├── CLAUDE.md # Claude Code
│ ├── .cursorrules # Cursor
│ ├── AGENTS.codex.md # Codex / OpenAI
│ ├── .windsurfrules # Windsurf / Antigravity
│ ├── copilot-instructions.md# GitHub Copilot
│ └── ollama-system.md # Ollama / local LLMs
├── ai-rules/
│ ├── context/ # YOUR project facts (fill once): project, file-map, data-model, api-map, env-keys
│ ├── rules/ # security · performance · database · frontend · error-handling · testing · code-quality
│ ├── skills/ # context-budget · debug-fix · tdd · test-writer · refactor · pr-reviewer · ship · setup
│ ├── agents/ # security · code · performance · frontend · docs reviewers
│ ├── hooks/ # session-start · pre-commit · pre-ship · post-error
│ ├── memory/ # rolling session summaries + indexed error ledger
│ └── scripts/ # session-save · error-log · secret-scan · backup · prune-sessions
├── plugins/ # OPTIONAL: Obsidian · Supabase MCP · notebooklm-py · minimal-context
└── examples/ # filled-in context examples (web app, mobile app, API)
Option 1 — Let your AI do it (easiest)
Open Claude Code, Cursor, or Codex in your project and send this:
Clone https://github.com/oreum-ai/smartstack into a temp folder, copy AGENTS.md and the ai-rules folder into my project root, then delete the temp folder.
Done. Your AI handles everything.
Option 2 — Download ZIP (no terminal needed)
- Click the green Code button at the top of this page
- Click Download ZIP and unzip it
- Copy everything inside the
smartstackfolder into your project root
Option 3 — Installer
git clone https://github.com/oreum-ai/smartstack
bash smartstack/install.shOption 4 — Manual
git clone https://github.com/oreum-ai/smartstack
cp -r smartstack/ai-rules smartstack/AGENTS.md your-project/Open ai-rules/context/project.md and fill in your project details. Takes 10 minutes, one time only. This is what stops your AI from forgetting your project every session.
Then open your AI tool and send this:
Read AGENTS.md and ai-rules/hooks/session-start.md, then summarize my project back to me in 5 bullets. Don't write any code yet.
If the bullets look right, you're set.
| Tool | File |
|---|---|
| Claude Code | adapters/CLAUDE.md → copy as CLAUDE.md in your project root |
| Cursor | adapters/.cursorrules → copy into your project root |
| Codex | Nothing extra — reads AGENTS.md automatically |
| Windsurf / Antigravity | adapters/.windsurfrules → copy into your project root |
| GitHub Copilot | adapters/copilot-instructions.md → copy into .github/ folder |
| Ollama / local models | adapters/ollama-system.md → paste into your system prompt |
If you used the installer, it already handled this step.
SmartStack's core works completely standalone. These add extra power if you want them — each has a setup
guide in /plugins:
- Obsidian vault — turn your
ai-rules/memoryinto a searchable knowledge base; save tokens by linking notes instead of re-pasting. - Supabase MCP — let the assistant query/inspect your database safely through the Model Context Protocol.
- notebooklm-py — generate audio/topic overviews of your codebase docs.
- minimal-context ("caveman") mode — an ultra-lean profile for small local models that strips everything to essentials.
| Without SmartStack | With SmartStack |
|---|---|
| Re-explain the stack every prompt | AGENTS.md is read once, automatically |
| Assistant rescans the repo to "understand" | It reads short context/ summaries by index |
| Same bug debugged from scratch again | debug-fix checks the error ledger first |
| Giant 10-task prompts that degrade | Hooks enforce focused 3–5 task sessions |
| Tool | Adapter file | Status |
|---|---|---|
| Claude Code | CLAUDE.md |
✅ auto-read |
| Cursor | .cursorrules + .cursor/rules/ |
✅ auto-read |
| Codex / OpenAI | AGENTS.md |
✅ auto-read |
| Windsurf / Antigravity | .windsurfrules |
✅ auto-read |
| GitHub Copilot | .github/copilot-instructions.md |
✅ auto-read |
| Ollama / local LLMs | ollama-system.md (paste into system prompt / Modelfile) |
✅ manual |
SmartStack is MIT licensed — use it commercially, fork it, build on it. Keep the copyright notice.
Created and maintained by oreumaii. See LICENSE and NOTICE.
If SmartStack helps you, a ⭐ on the repo and a "Powered by SmartStack" line in your README are appreciated.