AADITHYA VIMAL

AI · 2026 · Live demo · Open source · 2026

Fuzit

A local-first context engineering platform that bridges massive codebases and LLM context windows using typed dependency graphs, Knapsack token optimization, and zero-telemetry secret redaction.

My role
Designed and built the TypeScript monorepo, AST symbol extractors, knapsack token budget solver, MCP server, and VS Code extension.
Status
Live demo · Open source · 2026
Category
AI · Local-first context engineering
Live demoSource on GitHub

Live demo · interactive

🔒fuzit.pages.dev
CONNECTING TO FUZIT...

Embedded for convenience — if it doesn’t load here, open the full site directly.

01

Understand

Why it exists. Naive developer tools paste entire files or blind folder dumps into prompts, exceeding token budgets, driving up costs, and leaking private keys. Fuzit continuously synthesizes high-density, explainable context bundles 100% locally.

What I built. Modular TypeScript monorepo operating across 5 local surfaces: CLI (fuzit), one-watcher-per-root live daemon, stdio Model Context Protocol (MCP) server for Claude Desktop/Cursor/Windsurf, native VS Code extension, and an out-of-process restricted plugin host.

The problem it solves. Naive developer tools dump entire repositories or flat file concatenations into LLM prompts, blowing past token limits, leaking private credentials, and polluting context with irrelevant code.

What it does

  • Typed Context Graph connecting AST symbols, functions, classes, imports, and Git commit history
  • Multi-pass Knapsack budgeting enforcing strict token limits with priority decay
  • 5 native surfaces: CLI, live file daemon, stdio MCP Server, VS Code extension, and isolated plugin host
  • Zero telemetry and automatic regex/entropy redaction of API keys, tokens, and .env secrets
02

Technical depth

Architecture. Extracts AST syntax trees with Tree-sitter, models project evidence into a typed Context Graph (symbols, imports, Git commit metadata, structural relationships), and solves token allocation using knapsack budgeting with priority decay and AST symbol ranking across Markdown, XML, JSON, and Text formats.

Security model. 100% local-first with zero telemetry. Built-in entropy scanner and regex pattern matcher redacts API keys, credentials, and .env files before context rendering. Path containment guard prevents symlink escapes and directory traversal.

Tradeoff. Local-first AST parsing and graph neighbor expansion incur initial index computation, but guarantee that source code and secrets never touch external servers.

  • TypeScript
  • Tree-sitter
  • tiktoken
  • MCP Server
  • VS Code API
  • Turborepo
How it works, step by step

Extracts AST syntax trees with Tree-sitter, models project evidence into a typed Context Graph (symbols, imports, Git commit metadata, structural relationships), and solves token allocation using knapsack budgeting with priority decay and AST symbol ranking across Markdown, XML, JSON, and Text formats.

Security notes

100% local-first with zero telemetry. Built-in entropy scanner and regex pattern matcher redacts API keys, credentials, and .env files before context rendering. Path containment guard prevents symlink escapes and directory traversal.

AI