mcplens

mcplens

Semantic codebase search for AI coding assistants — 70-85% token reduction, 100% local, zero cloud dependency.

AI coding assistants like Claude Code, Cursor, and Codex are powerful — but they have a fundamental problem: when you ask a question, they read files by guessing which ones arerelevant based on path and filename heuristics. On a medium-sized project, a single query can consume 10,000–20,000 tokens of context just loading files that may not even berelevant.

claude-context-optimizer solves this by giving your AI assistant semantic search over your codebase. Instead of reading files blindly, it callssearch_code("how does payment work?") and gets back only the 5 most relevant code chunks — indexed locally using embeddings, stored in SQLite, zero data leaving your machine.

How it works

When you open your AI assistant in a project:

1. The MCP server starts automatically (spawned via stdio by the assistant)
2. It compares file hashes against the last index and re-indexes only what changed (delta indexing)
3. A file watcher keeps the index in sync as you code
4. Your assistant now has access to 3 semantic search tools instead of reading raw files

You ask: "how does the Asaas webhook work?"

Without cco:                          With cco:
  Read AsaasWebhookController.php       search_code("asaas webhook")
  Read AsaasWebhookService.php          → returns 5 relevant chunks
  Read PaymentService.php               → ~800 tokens total
  Read BillingModule.php
  Read ...8 more files
  → ~15,000 tokens total

Under the hood

• Embeddings:Ollama with nomic-embed-text (768-dim) — 100% local, free, no API key
• Vector store: SQLite with cosine similarity computed in-process — no extra infrastructure
• Chunking: AST-aware via tree-sitter (splits by function/class) with sliding window fallback
• Transport: MCP stdio — the assistant spawns the process and communicates via pipe
• Persistence: Index lives in .claude-context/index.db and survives between sessions

Compatibility

claude-context-optimizer works with any MCP-compatible AI coding assistant. MCP (Model Context Protocol) is an open standard — the same server works across all clientswithout modification.

The init command detects which assistants you use and registers the server automatically in the right place.

Token savings

The index lives locally. The assistant fetches only what's relevant. The numbers speak for themselves:

These are context tokens — the portion you control. Savings scale with project size because larger projects trigger more heuristic file reads by default.

Tools exposed

Add this to your project's CLAUDE.md (or equivalent) to guide the assistant:

## Context Search

Always use MCP tools before reading files:

- search_code() — for conceptual or natural language queries
- get_symbol() — for exact class/function/method lookups
  Only read full files if both tools return insufficient context.

Installation options

Option A — npm (requires Ollama)

Zero overhead. Best for developers who already have Ollama installed.

npm install -g @vmsfigueredo/mcplens
ollama pull nomic-embed-text:latest
cd your-project && mcplens init

See INSTALL.md for full setup instructions.

Option B — Docker

Not available yet. Docker distribution (bundling Node + Ollama + model) is planned but not implemented. Track progress in the Roadmap.

Configuration

.claude-context/config.json is created automatically by init. Edit it to customize behavior:

{
  "embeddings": {
    "provider": "ollama",
    "ollamaUrl": "http://localhost:11434",
    "ollamaModel": "nomic-embed-text:latest"
  },
  "search": {
    "topK": 5,
    "minScore": 0.3
  },
  "ignore": [
    "**/tests/fixtures/**"
  ]
}

To use OpenAI embeddings instead:

{
  "embeddings": {
    "provider": "openai",
    "openaiApiKey": "sk-...",
    "openaiModel": "text-embedding-3-small"
  }
}

What gets indexed

Included by default:.ts .tsx .js .jsx .mjs .php .svelte .vue .py .rb .go .rs .css .scss .json .yaml .yml .md .sql

Ignored by default:node_modules, .git, vendor, dist, build, .next, .claude-context

The .claude-context/ directory is automatically added to .gitignore.

Index size reference

Dashboard

A lightweight web dashboard is available at http://localhost:3000 while the server is running:

• Overview — files indexed, chunks, index size, Ollama status
• Activity — live feed of re-indexing events
• Search — test queries manually and see scores (useful for calibrating minScore)
• Files — full list of indexed files with chunk counts

The dashboard runs on port 3333 by default. If that port is already taken (e.g. two projects open simultaneously), the port is automatically calculated from the project name. Toopen:

mcplens dashboard

To disable: add --no-dashboard to the server args in your MCP config.

Privacy

Everything runs on your machine:

• Embeddings are generated locally via Ollama — your code never leaves
• The index is stored in .claude-context/index.db in your project
• No telemetry, no analytics, no accounts

⚠️ If you use the OpenAI embeddings option, chunks are sent to OpenAI's API.

Why not just use existing tools?

The goal is to be the most accessible option for JS/TS developers — not the most feature-complete. If you already have Node.js, you're one command away.

Roadmap

• AST-based chunking via tree-sitter
• Delta indexing by file hash
• Real-time file watcher
• Dashboard
• Multi-client init (Claude Code, Cursor, Windsurf, Trae)
• Hybrid search (BM25 + semantic)
• Docker option with bundled Ollama
• Contextual retrieval (LLM-generated chunk summaries)
• Token usage analytics via Claude Code hooks

Contributing

PRs welcome. See INSTALL.md for local development setup.

Built with

This project was built using Claude Code — which is exactly why it exists.

License

MIT