Context acquisition for AI agents: find curated library docs, search the open web, and read complete original sources.

ctx library docs react "useEffect cleanup"     # find doc sources via Context7 index
ctx web search "OpenAI latest model pricing" --type news
ctx read <url>                                  # read full document as clean markdown
ctx web crawl start <docs-site> --limit 20      # start a reusable crawl job

AI coding agents need documentation, but current tools make trade-offs:

• RAG-based tools (ctx7, etc.) return 60-200 token fragments — too small for real understanding
• Full-doc tools (ref, etc.) return complete pages but search accuracy is inconsistent

ctx combines curated Context7 library discovery with open-web EXA search, then fetches the full originals via GitHub API, HTTP content negotiation, or headless browser rendering.

curl -fsSL https://raw.githubusercontent.com/ethan-huo/ctx/main/install.sh | sh

The installer downloads the latest GitHub Release binary for your platform, verifies the checksum, and overwrites the existing ctx binary. Supported release targets:

• macOS ARM64
• Linux ARM64
• Linux AMD64

Install a specific release:

curl -fsSL https://raw.githubusercontent.com/ethan-huo/ctx/main/install.sh | sh -s -- --version v0.1.0

Install somewhere other than ~/.local/bin:

curl -fsSL https://raw.githubusercontent.com/ethan-huo/ctx/main/install.sh | sh -s -- --dir /usr/local/bin

make build    # compile to bin/ctx
make install  # build + symlink to ~/.local/bin/ctx

ctx library search mlx-swift "GPU stream thread safety"  # optional: discover IDs
ctx library docs mlx-swift "GPU stream thread safety"    # curated docs source map
ctx web search "GPU stream thread safety mlx swift" --type code
ctx read <url>                                             # read as clean markdown

Long documents (>2000 lines) automatically produce a structural summary with numbered sections:

ctx read <url>             # returns summary with section numbers
ctx read <url> -s 2.1      # read specific section
ctx read <url> -s "1-3,5"  # combine sections
ctx read <url> --toc       # compact heading outline

When plain HTTP isn't enough, ctx uses Cloudflare Browser Rendering for full JS-rendered pages:

All commands support -d for full API control (cookies, viewport, JS injection, etc.):

{
  url: "https://example.com",
  cookies: "session=abc",
  viewport: {width: 1920, height: 1080},
  addScriptTag: [{content: "document.querySelector('.nav')?.remove()"}]
}

Store headers that auto-inject into all requests for a domain:

ctx site set example.com Cookie "session=abc"
ctx site set example.com Authorization "Bearer token123"
ctx site ls

Issue reads auto-expand comments until a line budget is reached, then append a continuation hint like ctx read github://owner/repo/issues/123 --comments 9-20. Use --comments 1-3 or --comments all to override. GitHub directory reads return an ls-style listing headed by the canonical github://... path, so agents can immediately drill into a child path instead of trying to parse GitHub HTML. YouTube video URLs are rejected with a clear unsupported error. ctx does not provide transcript extraction because that path depends on unstable authenticated extraction and external downloader behavior.

ctx auth login ctx7          # Context7 (device code login, opens browser)
ctx auth login exa --key ... # EXA open-web search
ctx auth login cloudflare    # Cloudflare Browser Rendering
ctx auth status              # check what's configured

GitHub reads use your gh auth token automatically.

ctx ships with a skill definition for AI agents (Claude Code, Cursor, etc.) that teaches them the full search → read → navigate → scrape workflow. Install it with your agent's skill mechanism.