pikr — Pick the right context

Stop sending 100K tokens when 4K will do. Local RAG for AI coding tools — indexes your codebase and sends only the relevant code as context.

The problem

AI coding tools (Claude Code, GitHub Copilot, Cursor) burn through tokens by sending entire files as context — most of it irrelevant. There's no built-in way to selectively pick which functions or classes get sent.

pikr fixes this. It indexes your codebase locally and exposes a search_codebase MCP tool. Instead of reading full files, your AI agent retrieves only what's relevant — by semantic meaning, recency, and proximity to open files.

Result: ~4–8K focused tokens instead of 100K+

Features

• 🔒 100% local — embeddings, vector DB, and search all run on your machine. No data leaves.
• 🌳 AST-aware chunking — TypeScript & JavaScript parsed by function/class via tree-sitter, not by line count
• 🌐 Multi-language — regex-based chunking for Python, Go, Rust, PHP, Ruby, Swift, Kotlin, Java, C#
• ⚡ Local embeddings — all-MiniLM-L6-v2 via ONNX Runtime (~80 MB, downloaded once, runs offline)
• 🎯 Smart reranking — combines semantic similarity, recency decay, and open-file proximity
• 🔌 MCP server — Claude Code, Cursor, Windsurf, and any MCP-compatible agent can call search_codebase
• 📊 Dashboard — stats, settings, file browser, and 1-click MCP setup — all in the VS Code sidebar
• 💰 Token budget — configurable cap (default 4K) so you never overshoot

Install

From VS Code Marketplace — search for pikr in the Extensions panel, or:

ext install eksant.pikr

Quick Open (Ctrl+P / Cmd+P):

ext install eksant.pikr

From .vsix (manual):

code --install-extension pikr-0.1.0.vsix

Getting started

1. Extension activates — indexing starts automatically

On first activation, pikr downloads the embedding model (~80 MB) to ~/.pikr/models/. Watch progress in View → Output → pikr.

Indexing starts automatically for all supported files (excluding node_modules, dist, .git):

2. Open the dashboard

Click the pikr icon in the activity bar. Three tabs + one optional tab:

3. Connect to your AI agent (optional but recommended)

Open the MCP tab in the dashboard and click "Setup Claude Code MCP" — it runs the setup command in your terminal automatically.

Or manually:

claude mcp add pikr -- node /absolute/path/to/pikr/dist/mcp.js

Then verify:

claude mcp list
# → pikr — Connected

Commands

Open the Command Palette (Cmd+Shift+P / Ctrl+Shift+P):

MCP Integration

pikr ships a standalone MCP server (dist/mcp.js) with one tool:

Works with Claude Code, Cursor, Windsurf, Gemini CLI, and any stdio MCP-compatible agent.

Important: Run your AI agent from your project root so pikr finds the right index (~/.pikr/<hash-of-cwd>/). Open the project in VS Code first to ensure it's indexed.

Settings

How it works

VS Code Extension:
  File watcher → AST/regex chunker → ONNX embedder → LanceDB (~/.pikr/<hash>/)

MCP Server / Search command:
  Query → ONNX embedder → vector search → reranker → context builder → AI agent

Stack: TypeScript · LanceDB · ONNX Runtime · tree-sitter · MCP SDK · esbuild

• LanceDB — embedded vector DB, no server needed, same tech used by Continue IDE
• tree-sitter — pure-JS grammars for TS/JS (works in VS Code's Electron environment)
• all-MiniLM-L6-v2 — 384-dimension embeddings, fast enough for local use, accurate enough for code

Local development

git clone https://github.com/eksant/pikr
cd pikr
npm install
npm run build

Press F5 in VS Code to open the Extension Development Host with pikr loaded.

npm run watch        # rebuild on save
npm run lint         # ESLint
npm run format       # Prettier
npm test             # vitest unit tests

Requirements

• VS Code 1.85+
• Internet connection on first activation (one-time model download ~80 MB)

Issues & Feedback

Found a bug or have a feature request? Please open an issue:

→ github.com/eksant/pikr/issues

Questions or ideas? Start a discussion:

→ github.com/eksant/pikr/discussions

License

MIT © 2025 eksant