Shrink your AI context window without shrinking your codebase

What it does

codesight scans a repository and generates compact markdown context files—CODESIGHT.md, tool-specific configs, and a per-topic wiki—so AI assistants can load a 200-token index instead of grepping through dozens of files. It parses TypeScript via AST and falls back to regex for thirteen other languages, detecting routes, models, components, and dependencies without calling any external APIs.

The interesting bit

The wiki mode is the real savings engine. Rather than injecting a monolithic 5K-token map into every conversation, it slices the context into articles like auth.md or database.md, so the AI only pulls what it needs. A separate knowledge mode does the same for your Obsidian vault, ADRs, and meeting notes, giving the model access to “why we chose PostgreSQL” alongside “what the schema looks like.”

Key highlights

• Zero dependencies, fully offline, no API keys or configuration
• TypeScript gets AST-level precision; everything else relies on regex heuristics across 30+ frameworks
• Generates topical wiki articles (overview, auth, payments, UI) that persist across sessions
• Knowledge mode ingests markdown notes and classifies them as decisions, retros, specs, or meeting notes
• Exposes 13 MCP tools for Claude Code and Cursor, plus blast-radius analysis for individual files
• Benchmarks on production codebases claim 7× to 130× token reduction, though “exploration” savings are estimated from route/model counts rather than live API logs

Caveats

• Non-TypeScript languages rely on regex detection, so precision depends on how idiomatic your code is
• NoSQL schemas (e.g., MongoDB without an ORM) parse as zero models because there are no SQL declarations to extract
• The headline token-savings figures are projections based on formulas (routes × 400, models × 300), not measured API usage

Verdict

Grab it if you’re hitting context limits or paying per token in a large TypeScript project. Skip it if your codebase is small enough that a well-written README already covers the map.