codesight

Your AI assistant wastes thousands of tokens every conversation just figuring out your project. codesight fixes that in one command.

4,000+ downloads and counting.

Zero dependencies. AST precision. 30+ framework detectors. 14 ORM parsers. 14 MCP tools. One npx call.

Works with TypeScript, JavaScript, Python, Go, Ruby, Elixir, Java, Kotlin, Rust, PHP, Dart, Swift, C#, and BrightScript/BrighterScript ( Roku). TypeScript projects get full AST precision. Everything else uses battle-tested regex detection across the same 30+ frameworks.

Built by Kailesk Khumar, founder of HouseofMVPs and Kailxlabs

Also: ultraship (39 expert skills for Claude Code) · claude-rank (SEO/GEO/AEO plugin for Claude Code)

0 dependencies · Node.js >= 18 · 145 tests · 14 MCP tools · MIT · tested on 25+ OSS projects across 14 languages

Works With

Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Windsurf, Cline, Aider, and anything that reads markdown.

Related MCP server: repowise

Install

npx codesight

That's it. Run it in any project root. No config, no setup, no API keys.

npx codesight --wiki                       # Generate wiki knowledge base (.codesight/wiki/)
npx codesight --init                       # Generate CLAUDE.md, .cursorrules, codex.md, AGENTS.md
npx codesight --open                       # Open interactive HTML report in browser
npx codesight --mcp                        # Start as MCP server (14 tools) for Claude Code / Cursor
npx codesight --blast src/lib/db.ts        # Show blast radius for a file
npx codesight --profile claude-code        # Generate optimized config for a specific AI tool
npx codesight --benchmark                  # Show detailed token savings breakdown
npx codesight --native-ast                 # Opt-in: AST plugins for more languages (see docs/wasm-plugins.md)
npx codesight --mode knowledge             # Map knowledge base (.md notes → KNOWLEDGE.md)
npx codesight --mode knowledge ~/vault     # Map Obsidian vault, ADRs, meeting notes, retros

Wiki Knowledge Base (v1.6.2)

Inspired by Karpathy's LLM wiki pattern — but compiled from AST, not an LLM. Zero API calls. 200ms.

npx codesight --wiki

Generates .codesight/wiki/ — a persistent knowledge base of your codebase that survives across every session:

.codesight/wiki/
  index.md      — catalog of all articles (~200 tokens) — read this at session start
  overview.md   — architecture, subsystems, high-impact files (~500 tokens)
  auth.md       — auth routes, middleware, session flow
  payments.md   — payment routes, webhook handling, billing flow
  database.md   — all models, fields, relations, high-impact DB files
  users.md      — user management routes and related models
  ui.md         — UI components with props
  log.md        — append-only record of every wiki operation

Why this cuts token usage further:

Instead of loading the full 5K token context map every conversation, your AI reads one targeted article:

Persistent across sessions. The wiki lives in .codesight/wiki/, committed to git. Every new Claude Code, Cursor, or Codex session starts with full codebase knowledge from the first message.

Auto-regenerates. Use --watch to keep the wiki current as you code. Use --hook to regenerate on every commit.

3 new MCP tools for wiki access:

The key difference from general-purpose wiki tools: codesight already knows your routes, schema, blast radius, and middleware from AST — no LLM needed to extract code structure. The wiki is a narrative layer on top of data your codebase already contains.

Knowledge Mode (v1.9.3)

Not just code — your decisions, meeting notes, ADRs, and retrospectives carry as much context as the codebase itself. --mode knowledge maps them the same way codesight maps code.

npx codesight --mode knowledge              # Scan current directory for .md files
npx codesight --mode knowledge ~/vault      # Scan an Obsidian vault
npx codesight --mode knowledge ./docs       # Scan a project docs folder

Outputs .codesight/KNOWLEDGE.md — a compact AI context primer:

# Knowledge Map — my-project

> 47 notes · 12 decisions · 8 open questions · 2025-09-01 → 2026-04-01

## Key Decisions (12)

- [2026-03-20] Going with Polar.sh over Stripe Connect — simpler global payments
- [2026-03-15] Decided to use PostgreSQL — better JSON support and Drizzle compatibility
- [2026-02-10] Will use Redis for rate limiting — BullMQ already in stack

## Open Questions (8)

- Should we support PayPal later?
- When do we start the Stripe marketplace application?

## Note Index (47)

### Decision Records (8)

- `decisions/adr-002-payments.md` — 2026-03-20 — Going with Polar.sh over Stripe Connect
- `decisions/adr-001-database.md` — 2026-03-15 — We need a relational database...

### Meeting Notes (14)

### Retrospectives (6)

### Specs & PRDs (5)

### Research (4)

What it detects automatically:

Supports:

• Obsidian vaults (YAML frontmatter, [[backlinks]], #tags)

• Notion exports (.md files with frontmatter)

• ADR tooling (adr-tools, Log4brains, raw markdown)

• Any folder of markdown files

Used together:

Read .codesight/CODESIGHT.md   → what the code does
Read .codesight/KNOWLEDGE.md   → why decisions were made

CI: add npx codesight --mode knowledge alongside your existing codesight step. Both files stay current on every push.

Benchmarks (Real Projects)

Every number below comes from running codesight on real production codebases — both small SaaS projects (v1.6.2) and large open-source platforms with 4K–10K+ files (v1.6.4). Output tokens are measured from actual file size (chars / 4). Exploration tokens are estimated from what was extracted — routes × 400, models × 300, components × 250, etc. Route counts and model counts are cross-checked against actual source files.

Three-Level Token Reduction

codesight saves tokens at two distinct layers. The wiki (v1.6.2) adds a second layer on top of the base savings:

Average combined reduction: 91x. The wiki's "targeted" number = reading index.md at session start (~200 tokens) + one relevant article (~160-350 tokens depending on project). Your AI never loads the full context map for targeted questions.

The two savings layers are independent and compound:

Layer 1 — codesight scan eliminates manual file exploration. Instead of your AI running glob/grep/read across 40-138 files to understand the project, it reads one pre-compiled map.

Layer 2 — --wiki eliminates loading the full map for every question. Instead of loading 3K-5K tokens of full context at session start, your AI reads a 200-token index and pulls the one relevant article (~160-350 tokens) for each question.

Without codesight:   AI reads 26K-47K tokens per session exploring files
With codesight:      AI reads ~3K-5K tokens (the compiled map)
With --wiki:         AI reads ~200 tokens at start + ~300 per targeted question

Base Scan Results

SaaS C has 0 models because it uses MongoDB — no SQL ORM declarations for codesight to parse. This is correct detection, not a false negative.

Multi-Language OSS Benchmark (v1.6.7)

Tested against real open-source codebases spanning every supported language and framework. Output tokens are measured from actual file size. Exploration tokens are estimated (routes×400 + models×300 + components×250 + revisit multiplier). Zero false positives across all tests.

¹ Django project is GraphQL-first — 7 REST utility endpoints detected accurately, 0 false positives. ² High ratio on small boilerplate: Spring Boot route metadata compresses very well. ³ SvelteKit RealWorld app uses page routes (+page.svelte), not JSON API endpoints (+server.ts). 0 routes is correct.

How exploration tokens are estimated: routes×400 + models×300 + components×250 + hot_files×150 + env_vars×30, times a 1.3 revisit multiplier, minus the output size. This approximates what an AI would spend asking "what routes exist?", "show me the schema", etc. in a manual exploration session. Output token count is the actual measured file size.

Wiki Breakdown (v1.6.2)

"How does auth work?" — without wiki: loads 3,945 tokens. With wiki: reads auth.md (~350 tokens). 11x improvement per targeted question, 84x total vs manual.

Detection Accuracy

Verified against actual source files. Route counts cross-checked against route definitions; schema models cross-checked against ORM table declarations.

SaaS A's 5 missed routes use dynamic url.match(/pattern/) inside request handlers — a developer pattern that static analysis cannot resolve at scan time. This is an inherent limit of static analysis, not a framework gap. SaaS C missed an estimated 3 of 59 FastAPI routes. Zero false positives across all three projects.

Blast Radius Accuracy

Tested on a production SaaS: changing the database module correctly identified:

• 5 affected files across API, auth, and server layers

• All routes that touch the database

• 12 affected models (complete schema)

• BFS depth: 3 hops through the import graph

What Gets Detected

Measured across the three benchmark projects:

How It Works

codesight runs all 8 detectors in parallel, then writes the results as structured markdown. The output is designed to be read by an AI in a single file load.

What It Generates

.codesight/
  CODESIGHT.md     Combined context map (one file, full project understanding)
  routes.md        Every API route with method, path, params, and what it touches
  schema.md        Every database model with fields, types, keys, and relations
  components.md    Every UI component with its props
  libs.md          Every library export with function signatures
  config.md        Every env var (required vs default), config files, key deps
  middleware.md    Auth, rate limiting, CORS, validation, logging, error handlers
  graph.md         Which files import what and which break the most things if changed
  cicd.md          GitHub Actions / CircleCI pipelines (when present)
  githooks.md      lefthook / husky / raw .git/hooks (when present)
  skills.md        .claude/commands + .claude/skills (when present)
  report.html      Interactive visual dashboard (with --html or --open)

The last three come from built-in plugins that scan the dotfile directories (.github/, .husky/, .claude/) the main pass skips. They run automatically and stay silent on projects without those files — so they only ever add a section when there's something to report.

AST Precision

When TypeScript is installed in the project being scanned, codesight uses the actual TypeScript compiler API to parse your code structurally. No regex guessing.

AST detection is reported in the output:

Analyzing... done (AST: 60 routes, 18 models, 16 components)

No configuration needed. If TypeScript is in your node_modules, AST kicks in automatically. Works with npm, yarn, and pnpm (including strict mode). Falls back to regex for non-TypeScript projects or frameworks without AST support.

AST-supported frameworks: Express, Hono, Fastify, Koa, Elysia (route chains + middleware), NestJS (decorator combining + guards), tRPC ( router nesting + procedure types), Drizzle (field chains + relations), TypeORM (entity decorators), React (props from interfaces + destructuring + forwardRef/memo).

Native-AST WASM plugins (opt-in)

By default codesight uses its built-in extractors (AST for TypeScript, regex for everything else). Opt in with --native-ast to specify that WebAssembly plugins should be used to acquire full-AST precision for non-TypeScript source files:

npx codesight --native-ast                 # use every discovered plugin (additive)
npx codesight --native-ast=rust,go         # only these languages (authoritative for their files)
npx codesight --native-ast=none            # force off (overrides config)
npx codesight --native-ast-strict          # like --native-ast, but fail if a named plugin is missing
npx codesight --plugin-dir ./wasm          # extra directory to search for plugins

Dispatch to WASM plugins is language-driven: each plugin self-describes (via a describe() export) the file extensions it handles, so any language works — not just the ones with built-in detectors. Where no plugin handles a file, built-in extraction stands; with --native-ast unset, behavior is byte-identical to existing codesight behavior.

codesight's npm package ships no plugins — they are separate, opt-in artifacts. The project publishes prebuilt reference plugins (Rust/ syn, Python/ruff, Go/go/parser) as checksummed GitHub release assets; drop the ones you want to use into ~/.codesight/plugins/ (or point --plugin-dir at them), then enable them with --native-ast. Alternately, you may build your own plugin for any language against the contract.

→ Full contract, discovery rules, and download/install steps: docs/wasm-plugins.md.

Built-in plugins

Three first-party plugins run on every scan and surface context the main pass misses because it lives in skipped dotfile directories. Each is inert until its target files exist, so they cost nothing on projects that don't use them:

Opt out of any of them per project:

// codesight.config.js
export default {disableDetectors: ["cicd", "githooks", "skills"]};

Terraform is opt-in, not auto-loaded — it deliberately reaches outside the scanned directory (sibling ../infrastructure repos) and is most useful with an explicit service name, so it stays off until you ask for it:

// codesight.config.js
import {createTerraformPlugin} from "codesight/plugins/terraform";

export default {plugins: [createTerraformPlugin({infraPath: "../infra"})]};

Routes

Not just paths. Methods, URL parameters, what each route touches (auth, database, cache, payments, AI, email, queues), and where the handler lives. Detects routes across 30+ frameworks automatically.

Example output:

- `GET` `/api/users/me` [auth, db, cache]
- `PUT` `/api/users/me` [auth, db]
- `POST` `/api/projects` [auth, db, payment]
- `GET` `/api/projects/:id` params(id) [auth, db]
- `POST` `/webhooks/stripe` [db, payment]
- `GET` `/health`

Schema

Models, fields, types, primary keys, foreign keys, unique constraints, relations. Parsed directly from your ORM definitions via AST. No need to open migration files.

Example output:

### user

- id: text (pk)
- name: text (required)
- email: text (unique, required)
- role: text (default, required)
- stripeCustomerId: text (fk)

### project

- id: uuid (default, pk)
- ownerId: text (fk, required)
- name: text (required)
- settings: jsonb (required)
- _relations_: ownerId -> user.id

Dependency Graph

The files imported the most are the ones that break the most things when changed. codesight finds them and tells your AI to be careful.

Example output:

## Most Imported Files (change these carefully)

- `src/types/index.ts` — imported by **20** files
- `src/db/index.ts` — imported by **12** files
- `src/lib/auth.ts` — imported by **8** files
- `src/lib/cache.ts` — imported by **6** files
- `src/lib/env.ts` — imported by **5** files

Blast Radius

BFS through the import graph finds all transitively affected files, routes, models, and middleware.

npx codesight --blast src/db/index.ts

Example output:

  Blast Radius: src/db/index.ts
  Depth: 3 hops

  Affected files (10):
    src/api/users.ts
    src/api/projects.ts
    src/api/webhooks.ts
    src/auth/session.ts
    src/jobs/notifications.ts
    src/server.ts
    src/auth/index.ts
    src/jobs/cron.ts
    src/cli.ts
    src/index.ts

  Affected routes (33):
    GET /api/users/me — src/api/users.ts
    POST /api/projects — src/api/projects.ts
    POST /webhooks/stripe — src/api/webhooks.ts
    ...

  Affected models: user, session, account, project,
    subscription, notification, audit_log

Your AI can also query blast radius through the MCP server before making changes.

Environment Audit

Every env var across your codebase, flagged as required or has default, with the exact file where it is referenced.

Example output:

- `DATABASE_URL` **required** — .env.example
- `REDIS_URL` (has default) — .env.example
- `STRIPE_SECRET_KEY` **required** — src/lib/payments.ts
- `STRIPE_WEBHOOK_SECRET` **required** — .env.example
- `RESEND_API_KEY` **required** — .env.example
- `JWT_SECRET` **required** — src/lib/auth.ts

Token Benchmark

See exactly where your token savings come from:

npx codesight --benchmark

Example output (SaaS A — 138 files, Hono + Drizzle):

  Token Savings Breakdown:
  ┌──────────────────────────────────────────────────┐
  │ What codesight found         │ Exploration cost   │
  ├──────────────────────────────┼────────────────────┤
  │  38 routes                   │ ~15,200 tokens     │
  │  12 schema models            │ ~ 3,600 tokens     │
  │   0 components               │       0 tokens     │
  │  30 library files            │ ~ 6,000 tokens     │
  │  12 env vars                 │ ~ 1,200 tokens     │
  │   5 middleware               │ ~ 1,000 tokens     │
  │  20 hot files                │ ~ 3,000 tokens     │
  │ 138 files (search overhead)  │ ~11,040 tokens     │
  ├──────────────────────────────┼────────────────────┤
  │ codesight output             │ ~ 3,936 tokens     │
  │ Manual exploration (1.3x)    │ ~46,020 tokens     │
  │ SAVED PER CONVERSATION       │ ~42,084 tokens     │
  └──────────────────────────────┴────────────────────┘

How Token Savings Are Calculated

Each detector type maps to a measured token cost that an AI would spend to discover the same information manually:

The 1.3x multiplier accounts for AI revisiting files during multi-turn conversations. These estimates are conservative. A developer manually verified that Claude Code spends 40-70K tokens exploring the same projects that codesight summarizes in 3-5K tokens.

Roku / BrightScript / SceneGraph

codesight treats Roku channels as first-class projects. The manifest file at the channel root anchors detection — the same file Roku itself uses to identify a channel, so zero configuration is needed for the common case.

Standard single-channel layout (about 90% of Roku repos, matches the Roku docs' getting-started template and projects like rokucommunity/brighterscript-template):

/
  manifest
  source/         # Main.brs + shared .brs libraries
  components/     # *.xml + paired *.brs component handlers
  images/

codesight also recognizes the rokucommunity/brighterscript-template layout where the channel lives under src/ and the root carries a bsconfig.json for BrighterScript tooling.

Multi-channel monorepo layout (less common — used by larger codebases that ship several branded channels from one repo with roku-deploy + gulp to merge a shared common/ layer with per-channel assets at build time):

/
  package.json      # depends on roku-deploy, gulp
  gulpfile.js
  src/apps/
    common/         # shared layer, merged into every channel at build
    creatorA/
      manifest
    creatorB/
      manifest

This is detected via a strict structural signal: no manifest at root, roku-deploy in deps, and a common/ directory with at least 2 sibling directories that each have their own manifest. When the signal matches, each channel (plus common/) is registered as a workspace.

Mappings to codesight's data model

Configurable navigation helpers

Many Roku projects use a custom helper to switch the visible screen (names vary: ShowScreen, pushScreen, NavigateTo, showView, etc.). These are used as optional enrichment to tag routes as MODAL. Defaults cover the common conventions; override with rokuScreenHelpers in your codesight config if your project uses a different name:

{
  "rokuScreenHelpers": [
    "Router.push",
    "openScreen"
  ]
}

Routes are still detected from <children> even when no helper is present or when no call-site matches.

Example output

- `VIEW` `/homeView` — components/views/HomeView.xml
- `VIEW` `/detailView` — components/views/DetailView.xml
- `MODAL` `/errorModal` — components/modals/ErrorModal.xml

### DataTask

- requestUrl: string
- response: object

Supported Stacks

AI Config Generation

npx codesight --init

Generates ready-to-use instruction files for every major AI coding tool at once:

Each file is pre-filled with your project's stack, architecture, high-impact files, and required env vars. Your AI reads it on startup and starts with full context from the first message.

MCP Server (14 Tools)

npx codesight --mcp

Runs as a Model Context Protocol server. Claude Code and Cursor call it directly to get project context on demand.

{
  "mcpServers": {
    "codesight": {
      "command": "npx",
      "args": [
        "codesight",
        "--mcp"
      ]
    }
  }
}

OpenAI Codex CLI (~/.codex/config.toml):

[mcp_servers.codesight]
command = "npx"
args = ["codesight", "--mcp"]
startup_timeout_sec = 60

Codex timeout note: npx has to resolve the package on first run which can exceed the default 30-second timeout. Set startup_timeout_sec = 60 or install globally (npm install -g codesight) and use command = "codesight" instead — global installs start significantly faster.

Your AI asks for exactly what it needs instead of loading the entire context map. Session caching means the first call scans, subsequent calls return instantly.

AI Tool Profiles

npx codesight --profile claude-code
npx codesight --profile cursor
npx codesight --profile codex
npx codesight --profile copilot
npx codesight --profile windsurf

Generates an optimized config file for a specific AI tool. Each profile includes your project summary, stack info, high-impact files, required env vars, and tool-specific instructions on how to use codesight outputs. For Claude Code, this includes MCP tool usage instructions. For Cursor, it points to the right codesight files. Each profile writes to the correct file for that tool.

Visual Report

npx codesight --open

Opens an interactive HTML dashboard in your browser. Routes table with method badges and tags. Schema cards with fields and relations. Dependency hot files with impact bars. Env var audit. Token savings breakdown. Useful for onboarding or just seeing your project from above.

GitHub Action

Add to your CI pipeline to keep context fresh on every push:

name: codesight
on: [ push ]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 20
      - run: npm install -g codesight && codesight
      - uses: actions/upload-artifact@v4
        with:
          name: codesight
          path: .codesight/

Watch Mode and Git Hook

Watch mode re-scans automatically when your code changes:

npx codesight --watch

Only triggers on source and config files (.ts, .js, .py, .go, .prisma, .env, etc.). Ignores node_modules, build output, and non-code files. Shows which files changed before each re-scan. Your config (disabled detectors, plugins) is preserved across re-scans.

Git hook regenerates context on every commit:

npx codesight --hook

Context stays fresh without thinking about it.

All Options

npx codesight                              # Scan current directory
npx codesight ./my-project                 # Scan specific directory
npx codesight --wiki                       # Generate wiki knowledge base
npx codesight --init                       # Generate AI config files
npx codesight --open                       # Open visual HTML report
npx codesight --html                       # Generate HTML report without opening
npx codesight --mcp                        # Start MCP server (14 tools)
npx codesight --blast src/lib/db.ts        # Show blast radius for a file
npx codesight --profile claude-code        # Optimized config for specific tool
npx codesight --watch                      # Watch mode (add --wiki to auto-regenerate wiki)
npx codesight --wiki --watch               # Watch + auto-regenerate wiki on changes
npx codesight --hook                       # Install git pre-commit hook (includes wiki)
npx codesight --benchmark                  # Detailed token savings breakdown
npx codesight --json                       # Output as JSON
npx codesight --mode knowledge             # Map .md knowledge base → KNOWLEDGE.md
npx codesight --mode knowledge ~/vault     # Map Obsidian vault or any .md folder
npx codesight --max-tokens 50000           # Trim output to fit token budget
npx codesight --since HEAD~5               # Show routes from last 5 commits only
npx codesight -o .ai-context               # Custom output directory
npx codesight -d 5                         # Limit directory depth

How It Compares

codesight is purpose-built for the problem most developers actually have: giving their AI assistant enough context to be useful without wasting tokens on file exploration. It focuses on structured extraction (routes, schema, components, dependencies) rather than general-purpose code graph analysis.

Contributing

git clone https://github.com/Houseofmvps/codesight.git
cd codesight
pnpm install
pnpm dev              # Run locally
pnpm build            # Compile TypeScript
pnpm test             # Run 145 tests

PRs welcome. Open an issue first for large changes.

License

MIT

If codesight saves you tokens, star it on GitHub so others find it too.