v1.0.0 · MIT

Super Codebase
Context.

Auto-detects your project. Configures semantic search, architectural boundary enforcement, and a verify loop. Any language. Any team size.

bash

$ curl -fsSL https://raw.githubusercontent.com/MasoomChoudhury/super-codebase-context/main/install.sh | bash

✓ Detected: TypeScript · Next.js · pnpm

✓ .context/index.md generated

✓ CLAUDE.md generated

✓ Augment Context Engine MCP registered

✓ verify-tools MCP registered

✓ dependency-cruiser installed

✓ pre-commit hook installed

── Done in 87 seconds ──

copy install command

View on GitHub Quick Start →

< 2 min

Full stack install time

Architecture layers

70%

Agent perf improvement

API keys required

12+

Languages supported

Architecture

Seven layers, zero config

Each layer is independent and automatically configured based on what the installer detects in your project.

0Layer

Codebase Context Specification

.context/index.md .contextignore per-module stubs

The highest-ROI layer. Works like .editorconfig for AI tools — every agent reads this before touching your code, so it knows your architecture and module boundaries without re-explaining every session. Auto-generated from your detected project structure.

30 sec · highest ROI

1Layer

Agent Instructions

CLAUDE.md Understand-Anything Supermemory

Auto-generates CLAUDE.md with correct build commands and conventions. Understand-Anything builds a visual knowledge graph with /understand-diff for blast-radius analysis before every commit. Supermemory gives Claude persistent cross-session memory.

local · no cost

2Layer

Semantic Index (MCP)

Augment Context Engine search_codebase()

Free MCP server. No API keys, no cloud, no GPU. Semantic codebase indexing that answers "where does rate limiting happen?" even when no file contains that exact string. Benchmarked at 70% agent performance improvement across Claude Code, Cursor, and Codex.

free · no keys

3Layer

Verify Loop (MCP)

verify_syntax() run_tests() check_deps()

Custom local MCP server exposing your project's own toolchain as callable tools. Claude runs tsc --noEmit + linter + tests before presenting output — self-correcting instead of handing you broken code. Closes the generate → verify → self-correct loop.

local MCP

4Layer

Precision Search (Libraries)

retriv code-chunk

TypeScript-native hybrid BM25 + vector search (retriv) with AST-aware chunking that splits at function/class/method boundaries (code-chunk). Add when building a custom search pipeline. Up to 30% better recall than vector-only.

optional

5Layer

Custom Tooling Primitive

cased/kit tree-sitter

Production-grade codebase mapping, symbol extraction, and dependency analysis via tree-sitter (12+ languages). Use when building custom agents, code review bots, or documentation generators on top of your codebase.

if building

6Layer

Safety Net (CI + pre-commit)

dependency-cruiser pre-commit hook tsc linter

Architectural boundary enforcement. Define rules like "UI must not import from server." Pre-commit hook runs tsc + linter + dependency-cruiser automatically on every commit. Fails if an architectural boundary is crossed. Generate an SVG dependency graph with npm run check:deps:graph.

CI + git hook

MCP Tools

What Claude can call

These tools are registered in Claude Code's MCP config automatically. Claude calls them on its own — you don't have to ask.

Tool

What it runs

When called

search_codebase(query)

Semantic search across entire codebase via Augment Context Engine

Before reading any file — finds relevant code by meaning

verify_syntax()

Runs tsc --noEmit + linter, returns errors with line numbers

After every code generation step

run_tests(pattern?)

Runs test suite, optionally scoped to a file pattern

After touching test-covered code

check_deps()

Runs dependency-cruiser, returns architecture boundary violations

Before any refactor touching imports

/understand-diff

Builds knowledge graph diff — visualises blast radius of current changes

Manually before commits touching 3+ files

/understand-chat

Ask architecture questions in plain English over the knowledge graph

Anytime — "how does auth flow through this system?"

Output

Files created in your project

Everything committed to git so your entire team gets the same AI setup on git pull.

📋

.context/index.md

Architecture overview and module boundaries. Edit this first — 15 minutes here saves hours of AI confusion.

🤖

CLAUDE.md

Persistent Claude Code instructions. Auto-generated with your actual project name, languages, and build commands.

🏗️

.dependency-cruiser.cjs

Architectural boundary rules. Pre-configured with UI→server and lib→framework constraints. Tune to your folder structure.

🔒

.git/hooks/pre-commit

Auto-runs tsc + linter + dependency-cruiser on every commit. Fails if architecture is violated.

⚡

.scc/mcp-verify/

Local MCP server exposing verify_syntax, run_tests, check_deps as callable tools for Claude.

🚫

.contextignore

Excludes node_modules, dist, build, *.map from AI context. Like .gitignore for AI tools.

📁

src/components/.context/

Per-module context stubs. Auto-created for detected directories. Fill in module-specific conventions.

📦

package.json scripts

Adds check:deps and check:deps:graph (SVG dependency visualisation) to your existing package.json.

Compatibility

Works with every major AI coding agent

Context files, MCP servers, and pre-commit hooks are configured for all agents automatically.

Claude Code

✓ CLAUDE.md · MCP · Hook

Cursor

✓ .cursorrules · MCP · Hook

Codex CLI

✓ AGENTS.md · MCP · Hook

Windsurf

✓ Context · MCP · Hook

Gemini CLI

✓ Context · MCP · Hook

GitHub Copilot

~ copilot-instructions · Hook

Quick Start

Up and running in 5 steps

Run the installer

Run this in your project root. Works on macOS, Linux, and WSL on Windows. Requires Node.js ≥ 18 and git.

$ curl -fsSL https://raw.githubusercontent.com/MasoomChoudhury/super-codebase-context/main/install.sh | bash

Edit .context/index.md

This is the most important step. Fill in your actual architecture, module boundaries, and what must never be imported where. The auto-generated file is a scaffold — your domain knowledge is what makes it powerful.

Tune .dependency-cruiser.cjs

Match the from/to path patterns to your actual folder structure, then verify everything passes.

$ npm run check:deps
✓ No architectural violations found

Build the knowledge graph

Open Claude Code and run the Understand-Anything plugin. Builds once, persists forever. After this you can ask architecture questions in plain English.

# Inside Claude Code
/understand-anything
✓ Knowledge graph built (247 nodes, 891 edges)

Commit and share with your team

Every teammate gets the identical AI setup the moment they pull. New team member? One command and they're fully configured.

$ git add .context/ CLAUDE.md .dependency-cruiser.cjs .contextignore
$ git commit -m "chore: add codebase-setup AI stack"
$ git push

Super CodebaseContext.