Finally understand what changed in your dependencies—without leaving your terminal.

• No more tab juggling – One command aggregates releases, changelogs, and commits into a single summary
• Works with any repo – Even poorly documented projects (automatic fallback to commit history)
• Date or version ranges – "What changed since v3.0?" or "since last month?"
• CI/CD ready – JSON output, meaningful exit codes, post-update hooks
• AI-enhanced – Optional categorization of breaking changes, features, fixes, and security updates

# Check what's new in a GitHub repo
npx @whatsnew/cli@latest vercel/ai

# Check what changed since a version
whatsnew facebook/react --since v18.2.0

# Check what changed since a date
whatsnew vercel/next.js --since 2024-06

# npm
npm install -g @whatsnew/cli

# pnpm
pnpm add -g @whatsnew/cli

# bun
bun add -g @whatsnew/cli

# Or run directly with npx (use @latest to avoid caching issues)
npx @whatsnew/cli@latest vercel/ai

brew tap gustavovalverde/wnf https://github.com/gustavovalverde/wnf
brew install whatsnew

Download from GitHub Releases for your platform:

• whatsnew-darwin-arm64 (macOS Apple Silicon)
• whatsnew-darwin-x64 (macOS Intel)
• whatsnew-linux-x64 (Linux x64)
• whatsnew-linux-arm64 (Linux ARM64)
• whatsnew-windows-x64.exe (Windows)

# .github/workflows/changelog-check.yml
name: Changelog Check
on:
  pull_request:
    paths:
      - 'package.json'

jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Show dependency changes
        run: npx @whatsnew/cli@latest vercel/ai --since v3.0.0 --format markdown >> $GITHUB_STEP_SUMMARY

whatsnew is deterministic-first: it uses rule-based parsing before considering AI. This makes results fast, reproducible, and free to run.

GitHub Releases  →  CHANGELOG.md  →  Commit History  →  AI Enhancement
    (primary)        (secondary)       (tertiary)        (when needed)

1. Release notes – Parsed with format detection (Changesets, Keep-a-Changelog, Conventional Commits)
2. Changelog files – Falls back to CHANGELOG.md if release notes are sparse
3. Commit history – Always available as a last resort
4. AI enhancement – Optional fallback when deterministic parsing produces low-quality results

Each response includes a confidence score so you know how reliable the data is.

AI enhancement is optional and only activates when:

• An AI API key is configured (see Environment Variables)
• Deterministic parsing produces low-quality results

What triggers AI fallback:

Real-world examples where AI helps:

• Repos with free-form release notes (no consistent format)
• Legacy projects with inconsistent changelog styles
• Auto-generated notes that don't follow conventions
• Non-English changelogs that deterministic rules miss

Without AI configured: whatsnew still works perfectly—you just get deterministic results with lower confidence scores for edge cases.

whatsnew extracts and presents information from source repositories. When commit messages or release notes are low-quality, the output reflects this limitation.

Example of poor source commits:

fix
fix: lint
fix: sidebar

Produces unhelpful output:

Bug Fixes
  [*] lint
  [*] sidebar

This is not a bug—whatsnew accurately reflects what's in the source. Repositories following Conventional Commits or Keep a Changelog produce significantly better results.

The confidence score helps identify this: Lower confidence often indicates sparse or low-quality source data.

See ADR-001: Input Quality Limitations for detailed analysis and future mitigation strategies.

vercel/ai v4.0.0 → v4.1.0

BREAKING CHANGES
  • Removed deprecated `useCompletion` hook
  • Changed `streamText` return type

FEATURES
  • Added `useObject` hook for structured data
  • New `anthropic` provider

FIXES
  • Fixed memory leak in streaming responses

Confidence: 0.92 | Sources: GitHub Release, CHANGELOG.md

whatsnew vercel/ai --format json | jq '.changes.breaking'

whatsnew vercel/ai --format markdown >> CHANGELOG_SUMMARY.md

For programmatic access, whatsnew also provides an HTTP API:

# Latest release
curl https://api.whatsnew.dev/v1/repos/github/facebook/react/releases/latest/whats-new

# Since a version
curl https://api.whatsnew.dev/v1/repos/github/vercel/ai/releases/whats-new?since=v3.0.0

# Monorepo package filter
curl https://api.whatsnew.dev/v1/repos/github/vercel/ai/releases/latest/whats-new?package=@ai-sdk/openai

whatsnew is built as a monorepo. Use individual packages for programmatic access:

whatsnew supports multiple configuration methods. Priority order (highest first):

1. CLI flags (--github-token, --ai-key)
2. Environment variables
3. Config file (~/.config/whatsnew/config.json)

# Set your GitHub token (one-time)
whatsnew config set github_token ghp_xxxxxxxxxxxx

# Optional: Set AI API key for enhanced parsing
whatsnew config set ai.api_key sk-ant-xxxxxxxxxxxx

# View current configuration
whatsnew config list

Located at ~/.config/whatsnew/config.json (XDG-compliant):

{
  "github_token": "ghp_xxx",
  "ai": {
    "provider": "anthropic",
    "api_key": "sk-ant-xxx"
  }
}

whatsnew config set <key> <value>   # Set a value (validates tokens)
whatsnew config list                # Show current config (tokens masked)
whatsnew config path                # Show config file location
whatsnew config unset <key>         # Remove a value

Environment variables override config file values:

# GitHub API access
GITHUB_TOKEN=ghp_xxx              # Recommended: 5,000 req/hr vs 60 without

# Display options
NO_COLOR=1                        # Disable colored output

# AI Enhancement (optional - set ONE of these to enable)
ANTHROPIC_API_KEY=sk-ant-xxx      # Direct Anthropic API
OPENAI_API_KEY=sk-xxx             # Direct OpenAI API
AI_GATEWAY_API_KEY=xxx            # Vercel AI Gateway (multi-provider)
AI_PROVIDER=anthropic             # anthropic or openai (default: anthropic)

For one-off overrides without modifying config:

whatsnew vercel/ai --github-token ghp_xxx
whatsnew vercel/ai --ai-key sk-ant-xxx

Note: AI is auto-enabled when an API key is present. It only runs when deterministic parsing produces low-quality results (see When AI Kicks In).

# Prerequisites: Bun 1.3.1+, Node 24+
bun install

# Build all packages
bun run build

# Run tests
bun run test

# Type check
bun run typecheck

# Lint
bun run lint

• Data Flow – How multi-source aggregation works
• CLI Architecture – Command design and implementation
• WNF Specification – Machine-readable changelog format
• Implementation Roadmap – Development phases
• Architecture Decisions – ADRs documenting key technical decisions

MIT