Reuse

A personal codebase registry that lets AI assistants reference patterns across your projects.

Register your projects with descriptions, tags, and notable patterns. Any MCP-compatible AI assistant can then search your registry, read code from referenced projects, and adapt patterns for your current work.

The AI reads your code to understand the approach, then uses its own judgment in the current context. Sometimes it replicates the pattern, sometimes it adapts it, sometimes it says "I see how you did it there, but here's a better fit for this project." It never blindly copies.

Setup with AI (Recommended)

The easiest way to set up Reuse is to point your AI assistant at this repo and let it help you.

Claude Code:

Clone https://github.com/stonekey908/reuse and set it up for me.
Install dependencies, build it, link it globally, and add the MCP
server to my global Claude config.

Cursor / Windsurf / any MCP-compatible editor:

Clone https://github.com/stonekey908/reuse, install and build it,
then configure it as an MCP server in my editor settings.

The AI will handle cloning, building, and wiring up the MCP config for your specific environment. Once set up, you can immediately start registering projects and asking the AI to reference them.

Screenshots

CLI — One command to see everything Reuse can do

Register a project through the web UI

Rich project detail view with tags, patterns, and links

Full MCP tool suite — 9 tools your AI assistant can call

MCP in action — updating and querying project metadata

AI-assisted code search and reuse commentary

Analysis tab — two-level theme tree, collapsible per section

Manual Setup

Prerequisites

Node.js 18+ (node --version to check)
npm (npm --version to check)
Git (git --version to check)

1. Clone and build

git clone https://github.com/stonekey908/reuse.git
cd reuse
npm install
npm run build        # Builds TypeScript server + Vite frontend

2. Link the CLI globally

npm link

This makes the reuse command available in your terminal from any directory.

Verify it works:

reuse --help

You should see the list of commands (list, search, add, remove, tag, pattern, serve).

3. Configure the MCP server

The MCP server is what lets AI assistants talk to Reuse. Configuration depends on your client.

Claude Code (CLI / Desktop App)

Add to ~/.claude.json under the top-level mcpServers key:

{
  "mcpServers": {
    "reuse": {
      "command": "node",
      "args": ["/FULL/PATH/TO/reuse/dist/mcp/stdio.js"]
    }
  }
}

Replace /FULL/PATH/TO/reuse with the actual path where you cloned the repo (e.g., /Users/yourname/reuse).

Restart Claude Code for the MCP server to load.

Claude Desktop

Add to your Claude Desktop config (Settings > Developer > Edit Config):

{
  "mcpServers": {
    "reuse": {
      "command": "node",
      "args": ["/FULL/PATH/TO/reuse/dist/mcp/stdio.js"]
    }
  }
}

Restart Claude Desktop.

Cursor

Add to .cursor/mcp.json in your home directory or project:

{
  "mcpServers": {
    "reuse": {
      "command": "node",
      "args": ["/FULL/PATH/TO/reuse/dist/mcp/stdio.js"]
    }
  }
}

Windsurf

Add to your Windsurf MCP settings:

{
  "mcpServers": {
    "reuse": {
      "command": "node",
      "args": ["/FULL/PATH/TO/reuse/dist/mcp/stdio.js"]
    }
  }
}

Any other MCP-compatible client

Reuse uses stdio transport — the standard MCP protocol over stdin/stdout. Any client that supports MCP stdio servers can connect. The command is:

node /FULL/PATH/TO/reuse/dist/mcp/stdio.js

No environment variables or API keys required.

4. Verify it works

In your AI assistant, try:

"List my reuse projects"

If the MCP is connected, the AI will call list_projects and show your registry (empty at first).

Registering Projects

Three ways to add projects — use whichever fits your workflow.

Via AI (natural language)

Once the MCP is connected, just tell the AI:

"Register my-app from ~/projects/my-app. It's a React dashboard with real-time WebSocket data and JWT auth."

The AI will call register_project with the path, description, and tags it infers. It can also auto-detect the git remote.

"Find that wine app in my Documents and register it"

The AI will use find_local_project to locate the directory, then register it.

Automatic pattern extraction

When a project is registered without patterns, Reuse proactively nudges the AI toward extract_patterns — a tool that scouts the project (README, package.json, directory tree, representative source files, and user-facing screens) and returns a structured scouting report. The AI reads a handful of the suggested files, identifies 6–12 distinctive patterns, and saves them with update_project.

"Add ~/projects/graph-engine to reuse"

Under the hood:

1. register_project({ name: "graph-engine", projectPath: "~/projects/graph-engine" })
   → Registered. Nudge: "No patterns supplied — call extract_patterns next."

2. extract_patterns({ name: "graph-engine" })
   → Scouting report: README excerpt, deps, tree, suggestedFilesToRead,
     userFacingScreens (route entry points), 9-step instructions.

3. read_project_file x3–6 on the most interesting files.

4. update_project({ name: "graph-engine", patterns: { ... } })
   → 6–12 named kebab-case patterns with capability tags + file evidence.

The scout deliberately surfaces both file-named signals (large meaningful modules in lib/, services/, hooks/, agents/, context/) AND capability anchors (every screen the user can navigate to). The 9-step instructions teach the AI to do a capability-walk — for every screen, ask "what reusable mechanism makes this work?" — so cross-file flows (image-pick → compress → multimodal-CF-proxy) get named even when no single file shouts "image-upload pipeline." A built-in NOT-A-PATTERN filter excludes design docs, HTML mockups, and PRDs.

Structured pattern shape

Patterns can be saved as plain strings or as structured objects. The structured shape lets the analysis pipeline cluster across projects by capability:

{
  "patterns": {
    "chunked-upload-with-retry": {
      "description": "Splits files into 2MB chunks, retries failed chunks 3 times with backoff. /src/upload/chunked.ts.",
      "capability": "chunked-upload-with-retry",
      "abstractionLevel": "feature",
      "domain": "frontend-web",
      "fileEvidence": ["/src/upload/chunked.ts"]
    }
  }
}

capability — kebab-case slug for the reusable idea (two projects with the same capability cluster together)
abstractionLevel — primitive | feature | discipline | architecture | spec
domain — frontend-mobile | frontend-web | frontend-desktop | backend-api | ai-integration | design-system | dev-tooling | distribution | infra-system | docs-content | build-tooling | testing-discipline
fileEvidence — exact file paths so the description stays grounded

A bare string description still works for back-compat — the registry auto-upgrades on read.

Via CLI

# Basic registration
reuse add my-app ~/projects/my-app

# With metadata
reuse add my-app ~/projects/my-app \
  -d "React dashboard with real-time data" \
  -t "react,typescript,websockets,auth"

# Add patterns after registration
reuse pattern my-app auth "JWT with refresh tokens and role-based access"
reuse pattern my-app real-time "WebSocket-based live data feeds with reconnection"

# Add tags
reuse tag my-app nextjs supabase

# Search
reuse search auth
reuse search websocket

Via Web UI

reuse serve
# Open http://localhost:3210

Click "+ Add Project" to register, or click "Edit" on any project to update all fields including path, git URL, patterns, and links. Changes are instant — the same ~/.reuse/registry.json is used by all three interfaces.

Using Reuse

Once projects are registered, you use Reuse by talking to your AI assistant naturally:

What you say	What the AI does
"I want file upload like my photos app"	Searches projects for "upload", reads the relevant code, adapts the pattern
"Show me how project-a handles encryption"	Gets project details, searches for encryption code, reads the files
"Build auth like project-b but for a mobile app"	Reads the auth pattern from project-b, adapts it for React Native
"What projects use Supabase?"	Searches tags for "supabase", lists matching projects
"Register that wine app from my Documents"	Finds the folder, detects git remote, registers it

The AI maintains full autonomy. It reads your code to understand the approach, then decides how to apply it. It might:

Replicate the pattern closely if the stack matches
Adapt it significantly for a different framework
Say "I see the approach but there's a better way for this use case"

MCP Tools Reference

Read tools (for referencing code)

Tool	Parameters	Description
`list_projects`	—	Browse all registered projects with descriptions, tags, and pattern names
`search_projects`	`query`	Search by keyword across names, descriptions, tags, and patterns
`get_project_details`	`name`	Full details for a project including file structure overview
`search_project_code`	`name`, `pattern`, `fileGlob?`	Search source code within a project (case-insensitive, regex supported)
`read_project_file`	`name`, `filePath`, `startLine?`, `endLine?`	Read a file or specific line range (large files return first 500 lines with pagination)

Write tools (for managing the registry)

Tool	Parameters	Description
`register_project`	`name`, `projectPath`, `description?`, `tags?`, `patterns?`, `git?`, `links?`	Register a new project (auto-detects git remote). Patterns accept either string descriptions or structured objects with `capability`/`abstractionLevel`/`domain`/`fileEvidence` tags. If no patterns are supplied, the response nudges the AI to run `extract_patterns` next.
`extract_patterns`	`name`	Scout a registered project for reusable patterns. Returns a structured report (README excerpt, package.json, directory tree, scored `suggestedFilesToRead`, `userFacingScreens`, 9-step capability-walk instructions) for the AI to identify 6–12 distinctive patterns.
`update_project`	`name`, `description?`, `tags?`, `patterns?`, `git?`, `links?`	Update any metadata field. Patterns merge (existing keys overwritten). String patterns preserve any existing tags on the key (description-only update); structured patterns replace tags wholesale.
`remove_project`	`name`	Unregister a project (does NOT delete files)
`find_local_project`	`name`, `searchIn?`	Search filesystem for a project folder by name

CLI Reference

Command	Description
`reuse list`	List all registered projects with metadata
`reuse search <query>`	Search projects by keyword
`reuse add <name> <path>`	Register a project (`-d` description, `-t` tags, `-g` git URL)
`reuse remove <name>`	Unregister a project (no files deleted)
`reuse tag <name> <tags...>`	Add tags to a project
`reuse pattern <name> <key> <desc>`	Add or update a named pattern
`reuse serve [-p port]`	Start the web UI (default port 3210)

Web UI

Run reuse serve and open http://localhost:3210.

View all registered projects at a glance
Add new projects with the form
Edit any field — path, description, tags, git URL, patterns, links
Remove projects
The Analysis tab clusters patterns across the whole registry into a two-level tree — top-level theme (AI & LLM, UI components, Data storage, Realtime, Image & media, Dev tooling, Distribution & CLI, Testing, Docs & spec, Observability, …) → cluster or standalone → pattern members. Sections are collapsible per-theme with an Expand all / Collapse all toggle, and a staleness banner lights up when projects change after the cached run
All changes write to the same registry file used by the CLI and MCP

Analysis & Evals

Reuse clusters patterns across all your registered projects by capability and surfaces consolidation opportunities. The analysis runs against any configured provider (Anthropic / OpenAI / Gemini / Ollama) — set ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, or OLLAMA_BASE_URL in .env.local and pick a provider/model in the web UI. Results are cached to the registry and re-run when patterns change.

# Analysis
reuse analyze              # cached if fresh, otherwise runs against the default provider
reuse analyze --refresh    # force a re-run

The web UI exposes the same analysis through a provider/model picker, with a Stop button, live elapsed timer, and a staleness banner that lights up when projects change after the cached run.

Two-level theme grouping

Analysis output is organised as a tree, not a flat list:

Theme (top level) — one of 12 canonical functional areas plus Other: ai-llm, ui-components, data-storage-sync, cloud-backend, state-background, realtime-messaging, image-media, dev-tooling, distribution-cli, testing-quality, docs-spec, observability-errors. Themes describe functional capability (what the code does); they're a different axis from the per-pattern domain field which describes implementation layer (where the code lives).
Capability cluster or standalone (second level) — human-readable cluster names like "Document upload", "Reusable modal shell", "Multi-provider AI routing".
Pattern members — the actual entries from your projects.

Each theme is a collapsible section in the web UI with an item count. Use the Expand all / Collapse all toggle to flip every section in one click.

The theme list was empirically grounded by analysing a real 92-pattern multi-app registry and cross-referenced against awesome-nodejs (~50 cats), awesome-react (~22), the CNCF Cloud Native Landscape (6 layers), and full-stack tech-stack categorisations. 12 themes turned out to be the sweet spot — fine enough that "where would I look for image processing?" has an obvious answer, coarse enough that the page doesn't degenerate into a flat scroll.

Built-in evals

Reuse ships three independent eval harnesses that measure different parts of the pipeline. They're independent — the scout eval is purely about how well the MCP populates the registry; the analysis evals are about how well the clustering reads the populated registry.

Eval	What it measures	Run
E0 — scout	Does the MCP scout (`extract_patterns`) give the AI enough signal to name the right patterns? Coverage of hand-curated ground-truth capabilities, precision of proposals, design-doc noise rate.	`node scripts/scout-eval.mjs`
E1 — analysis snapshot	Layer 1 — does the analysis pipeline produce a cluster set close to a known-good fixture?	`reuse eval`
E2 — analysis quality (LLM judge)	Layer 2 — Opus 4.7 grades the clustering on coherence, similarity quality, difference quality, consolidation usefulness, granularity. Writes a markdown report to `eval-results/`.	`reuse eval --quality`

Scout eval (E0) is the newest addition. It runs Opus 4.7 against four ground-truth projects, asks it to extract patterns from the scout report alone (no file reads), and judges the proposals against hand-written capability lists. The current target gates at 100% / 98% / 0% (coverage / precision / design-doc-noise) — useful for catching regressions in scout heuristics or SCOUT_INSTRUCTIONS tuning.

=== SUMMARY ===
project         coverage   precision  noise      proposed/gt
project-a       100%       100%       0%         12/8
project-b       100%       92%        0%         12/10
project-c       100%       100%       0%         12/9
project-d       100%       100%       0%         14/10
average         100%       98%

See docs/EVALS.md for the rubric, how to read judge reports, and the prompt-tuning workflow.

How It Works

You register:
  "project-a" → path, description, tags, notable patterns, git, links

AI receives prompt:
  "I want file upload like my photos app"

AI calls Reuse MCP tools:
  1. search_projects("file upload") → finds photos-app
  2. get_project_details("photos-app") → sees the upload pattern description
  3. search_project_code("photos-app", "upload") → finds relevant source files
  4. read_project_file("photos-app", "src/Upload/index.tsx") → reads the actual code
  5. Adapts the approach for the current project's stack and context

And during registration:

You register a new project:
  "Add ~/projects/graph-engine to reuse"

AI calls Reuse MCP tools:
  1. register_project(...) → registered, nudged to extract patterns
  2. extract_patterns("graph-engine") → scouting report with suggested files
  3. read_project_file x3–6 on the key files
  4. update_project({ name, patterns: {...} }) → 5–8 named patterns saved

The AI can also use your other tools alongside Reuse. If a project has a Linear link, the AI can check Linear for related tickets. If it has a Notion link, it can pull docs. Reuse is the index — your other tools provide the depth.

Registry Format

Projects are stored in ~/.reuse/registry.json:

{
  "projects": {
    "project-a": {
      "path": "/Users/you/project-a",
      "description": "Mobile messaging app",
      "tags": ["react-native", "expo", "firebase", "encryption"],
      "patterns": {
        "e2e-encryption": {
          "description": "E2E encryption using libsodium for messages and attachments. /lib/crypto.ts.",
          "capability": "encrypted-local-storage",
          "abstractionLevel": "primitive",
          "domain": "frontend-mobile",
          "fileEvidence": ["/lib/crypto.ts", "/lib/storage.ts"]
        },
        "file-upload": "Chunked upload with progress tracking, retry, and compression"
      },
      "git": "https://github.com/you/project-a",
      "links": {
        "linear": "https://linear.app/team/project/project-a",
        "figma": "https://figma.com/file/abc123"
      }
    }
  }
}

The two pattern shapes (string and structured object) coexist. Tagged patterns cluster across projects by capability; bare strings still work for back-compat.

You can edit this file directly if you prefer — it's just JSON.

Security

Scoped access — the MCP server can only read files from explicitly registered projects, never arbitrary paths
Path traversal blocked — resolved paths must stay within the project directory
Smart pagination — large files (>500 lines) return the first 500 lines with startLine/endLine support for reading specific sections
Read-only project files — the registry is read/write, but your actual project code is read-only
No network — Reuse never phones home, calls APIs, or sends data anywhere. It's entirely local.
No credentials — no API keys, tokens, or accounts needed

Troubleshooting

MCP server not connecting:

Check the path in your MCP config points to the actual dist/mcp/stdio.js file
Make sure you ran npm run build (or npm run build:server) after cloning
Restart your AI client after changing the MCP config

reuse command not found:

Run npm link from the reuse directory
Or use node /path/to/reuse/dist/cli/index.js directly

reuse serve fails with invalid choice: 'serve':

Debian/Ubuntu (and WSL) ship a reuse package (the SPDX license tool) at /usr/bin/reuse that shadows this CLI. Check with which -a reuse.

Quickest fix — add an alias to ~/.bashrc (or ~/.zshrc):

alias reuse='node /FULL/PATH/TO/reuse/dist/cli/index.js'

Or run the CLI directly: node /FULL/PATH/TO/reuse/dist/cli/index.js serve
Or uninstall the SPDX tool if you don't use it: sudo apt remove reuse

Search returning no results:

Reuse searches with a built-in Node.js grep (no external tools required)
If ripgrep (rg) is installed at /opt/homebrew/bin/rg or /usr/local/bin/rg, it will use that instead for better performance
Check the project path is correct: reuse list

Web UI won't start:

Make sure you built the frontend: npx vite build (from the reuse directory)
Check nothing else is using port 3210: reuse serve -p 3211

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 61 Commits
docs		docs
eval-results		eval-results
scripts		scripts
src		src
tests		tests
.env.example		.env.example
.gitignore		.gitignore
AUTOMATE-LOG.md		AUTOMATE-LOG.md
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json
tsconfig.node.json		tsconfig.node.json
vite.config.ts		vite.config.ts

Folders and files

Latest commit

History

Repository files navigation

Reuse

Setup with AI (Recommended)

Screenshots

CLI — One command to see everything Reuse can do

Register a project through the web UI

Rich project detail view with tags, patterns, and links

Full MCP tool suite — 9 tools your AI assistant can call

MCP in action — updating and querying project metadata

AI-assisted code search and reuse commentary

Analysis tab — two-level theme tree, collapsible per section

Manual Setup

Prerequisites

1. Clone and build

2. Link the CLI globally

3. Configure the MCP server

Claude Code (CLI / Desktop App)

Claude Desktop

Cursor

Windsurf

Any other MCP-compatible client

4. Verify it works

Registering Projects

Via AI (natural language)

Automatic pattern extraction

Structured pattern shape

Via CLI

Via Web UI

Using Reuse

MCP Tools Reference

Read tools (for referencing code)

Write tools (for managing the registry)

CLI Reference

Web UI

Analysis & Evals

Two-level theme grouping

Built-in evals

How It Works

Registry Format

Security

Troubleshooting

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages