Fast CLI for summarizing anything you can point at:

• Web pages (article extraction; Firecrawl fallback if sites block agents)
• YouTube links (best-effort transcripts, optional Apify fallback)
• Podcast RSS feeds (best-effort: transcribes latest enclosure via Whisper when configured)
• Remote files (PDFs/images/audio/video via URL — downloaded and forwarded to the model)
• Local files (PDFs/images/audio/video/text — forwarded or inlined; support depends on provider/model)

It streams output by default on TTY and renders Markdown to ANSI (via markdansi). At the end it prints a single “Finished in …” line with timing, token usage, and a best-effort cost estimate (when pricing is available).

Requires Node 22+.

• npx (no install):

npx -y @steipete/summarize "https://example.com"

• npm (global install):

npm i -g @steipete/summarize

• Homebrew (custom tap):

brew install steipete/tap/summarize

Apple Silicon only (arm64).

summarize "https://example.com"

Input can be a URL or a local file path:

npx -y @steipete/summarize "/path/to/file.pdf" --model google/gemini-3-flash-preview
npx -y @steipete/summarize "/path/to/image.jpeg" --model google/gemini-3-flash-preview

Remote file URLs work the same (best-effort; the file is downloaded and passed to the model):

npx -y @steipete/summarize "https://example.com/report.pdf" --model google/gemini-3-flash-preview

YouTube (supports youtube.com and youtu.be):

npx -y @steipete/summarize "https://youtu.be/dQw4w9WgXcQ" --youtube auto

Podcast RSS feed (transcribes latest episode enclosure):

npx -y @steipete/summarize "https://feeds.npr.org/500005/podcast.xml"

Apple Podcasts episode page (extracts stream URL, transcribes via Whisper):

npx -y @steipete/summarize "https://podcasts.apple.com/us/podcast/2424-jelly-roll/id360084272?i=1000740717432"

Spotify episode page (best-effort; resolves to full episode via iTunes/RSS enclosure when available — not preview clips; may fail for Spotify-exclusive shows):

npx -y @steipete/summarize "https://open.spotify.com/episode/5auotqWAXhhKyb9ymCuBJY"

This is “best effort” and depends on what your selected model/provider accepts. In practice these usually work well:

• text/* and common structured text (.txt, .md, .json, .yaml, .xml, …)
  • text-like files are inlined into the prompt (instead of attached as a file part) for better provider compatibility
• PDFs: application/pdf (provider support varies; Google is the most reliable in this CLI right now)
• Images: image/jpeg, image/png, image/webp, image/gif
• Audio/Video: audio/*, video/* (when supported by the model)

Notes:

• If a provider rejects a media type, the CLI fails fast with a friendly message (no “mystery stack traces”).
• xAI models currently don’t support attaching generic files (like PDFs) via the AI SDK; use a Google/OpenAI/Anthropic model for those.

Use “gateway-style” ids: <provider>/<model>.

Examples:

• openai/gpt-5-mini
• anthropic/claude-sonnet-4-5
• xai/grok-4-fast-non-reasoning
• google/gemini-3-flash-preview
• zai/glm-4.7
• openrouter/openai/gpt-5-mini (force OpenRouter)

Note: some models/providers don’t support streaming or certain file media types. When that happens, the CLI prints a friendly error (or auto-disables streaming for that model when supported by the provider).

--length controls how much output we ask for (guideline), not a hard truncation.

npx -y @steipete/summarize "https://example.com" --length long
npx -y @steipete/summarize "https://example.com" --length 20k

• Presets: short|medium|long|xl|xxl
• Character targets: 1500, 20k, 20000
• Optional hard cap: --max-output-tokens <count> (e.g. 2000, 2k)
  • Provider/model APIs still enforce their own maximum output limits.
  • If omitted, no max token parameter is sent (provider default).
  • Prefer --length unless you need a hard cap (some providers count “reasoning” into the cap).
• Minimums: --length numeric values must be ≥ 50 chars; --max-output-tokens must be ≥ 16.

• Text inputs over 10 MB are rejected before tokenization.
• Text prompts are preflighted against the model’s input limit (LiteLLM catalog), using a GPT tokenizer.

npx -y @steipete/summarize <input> [flags]

• --model <provider/model>: which model to use (defaults to auto)
• --model auto: automatic model selection + fallback (default)
• --model <name>: use a config-defined model (see “Configuration”)
• --language <lang> / --lang <lang>: output language (default auto = same as source content)
• --timeout <duration>: 30s, 2m, 5000ms (default 2m)
• --retries <count>: LLM retry attempts on timeout (default 1)
• --length short|medium|long|xl|xxl|<chars>
• --language, --lang <language>: output language (auto = match source; or en, de, english, german, ...)
• --max-output-tokens <count>: hard cap for LLM output tokens (optional; only sent when set)
• --cli [provider]: use a CLI provider (case-insensitive; equivalent to --model cli/<provider>). If omitted, uses auto selection with CLI enabled.
• --stream auto|on|off: stream LLM output (auto = TTY only; disabled in --json mode)
• --render auto|md-live|md|plain: Markdown rendering (auto = best default for TTY)
• --format md|text: website/file content format (default text)
• --preprocess off|auto|always: controls uvx markitdown usage (default auto; always forces file preprocessing)
  • Install uvx: brew install uv (or https://astral.sh/uv/)
• --extract: print extracted content and exit (no summary) — only for URLs
  • Deprecated alias: --extract-only
• --json: machine-readable output with diagnostics, prompt, metrics, and optional summary
• --verbose: debug/diagnostics on stderr
• --metrics off|on|detailed: metrics output (default on; detailed adds a compact 2nd-line breakdown on stderr)

--model auto builds candidate attempts from built-in rules (or your model.rules overrides). CLI tools are not used in auto mode unless you explicitly enable them via cli.enabled in config. Why: CLI adds ~4s latency per attempt and higher variance. Shortcut: --cli (with no provider) uses auto selection with CLI enabled.

When enabled, auto prepends CLI attempts in the order listed in cli.enabled (recommended: ["gemini"]), then tries the native provider candidates (with OpenRouter fallbacks when configured).

Enable CLI attempts:

{
  "cli": { "enabled": ["gemini"] }
}

Disable CLI attempts:

{
  "cli": { "enabled": [] }
}

Note: when cli.enabled is set, it’s also an allowlist for explicit --cli / --model cli/....

Non-YouTube URLs go through a “fetch → extract” pipeline. When the direct fetch/extraction is blocked or too thin, --firecrawl auto can fall back to Firecrawl (if configured).

• --firecrawl off|auto|always (default auto)
• --extract --format md|text (default text)
• --markdown-mode off|auto|llm (default auto; only affects --format md for non-YouTube URLs)
  • auto: use an LLM converter when configured; may fall back to uvx markitdown
  • llm: force LLM conversion (requires a configured model key)
  • off: disable LLM conversion (still may return Firecrawl Markdown when configured)
• Plain-text mode: use --format text.

--youtube auto tries best-effort web transcript endpoints first. When captions aren't available, it falls back to:

1. Apify (if APIFY_API_TOKEN is set): Uses a scraping actor (faVsWy9VTSNVIhWpR)
2. yt-dlp + Whisper (if YT_DLP_PATH is set): Downloads audio via yt-dlp, transcribes with local whisper.cpp when installed (preferred), otherwise falls back to OpenAI (OPENAI_API_KEY) or FAL (FAL_KEY)

Environment variables for yt-dlp mode:

• YT_DLP_PATH - path to yt-dlp binary
• SUMMARIZE_WHISPER_CPP_MODEL_PATH - optional override for the local whisper.cpp model file
• SUMMARIZE_WHISPER_CPP_BINARY - optional override for the local binary (default: whisper-cli)
• SUMMARIZE_DISABLE_LOCAL_WHISPER_CPP=1 - disable local whisper.cpp (force remote)
• OPENAI_API_KEY - OpenAI Whisper transcription
• FAL_KEY - FAL AI Whisper fallback

Apify costs money but tends to be more reliable when captions exist.

--video-mode transcript forces audio/video inputs (local files or direct media URLs) through Whisper first, then summarizes the transcript text. Prefers local whisper.cpp when available; otherwise requires OPENAI_API_KEY or FAL_KEY.

Single config location:

• ~/.summarize/config.json

Supported keys today:

{
  "model": { "id": "openai/gpt-5-mini" }
}

Shorthand (equivalent):

{
  "model": "openai/gpt-5-mini"
}

Also supported:

• model: { "mode": "auto" } (automatic model selection + fallback; see docs/model-auto.md)
• model.rules (customize candidates / ordering)
• models (define presets selectable via --model <preset>)
• media.videoMode: "auto"|"transcript"|"understand"
• openai.useChatCompletions: true (force OpenAI-compatible chat completions)

Note: the config is parsed leniently (JSON5), but comments are not allowed. Unknown keys are ignored.

Precedence:

1. --model
2. SUMMARIZE_MODEL
3. ~/.summarize/config.json
4. default (auto)

Set the key matching your chosen --model:

• OPENAI_API_KEY (for openai/...)
• ANTHROPIC_API_KEY (for anthropic/...)
• XAI_API_KEY (for xai/...)
• Z_AI_API_KEY (for zai/...; supports ZAI_API_KEY alias)
• GEMINI_API_KEY (for google/...)
  • also accepts GOOGLE_GENERATIVE_AI_API_KEY and GOOGLE_API_KEY as aliases

OpenAI-compatible chat completions toggle:

• OPENAI_USE_CHAT_COMPLETIONS=1 (or set openai.useChatCompletions in config)

OpenRouter (OpenAI-compatible):

• Set OPENROUTER_API_KEY=...
• Prefer forcing OpenRouter per model id: --model openrouter/<author>/<slug> (e.g. openrouter/meta-llama/llama-3.1-8b-instruct:free)
• Built-in preset: --model free (uses a default set of OpenRouter :free models).

Quick start: make free the default (keep auto available)

# writes ~/.summarize/config.json (models.free) and sets model="free"
summarize refresh-free --set-default

# now this defaults to free models
summarize "https://example.com"

# whenever you want best quality instead
summarize "https://example.com" --model auto

Regenerates the free preset (writes models.free into ~/.summarize/config.json) by:

• Fetching OpenRouter /models, filtering :free
• Skipping models that look very small (<27B by default) based on the model id/name (best-effort heuristic)
• Testing which ones return non-empty text (concurrency 4, timeout 10s)
• Picking a mix of “smart-ish” (bigger context_length / output cap) and fast models
• Refining timings for the final selection and writing the sorted list back

If --model free stops working (rate limits, allowed-provider restrictions, models removed), run:

summarize refresh-free

Flags:

• --runs 2 (default): extra timing runs per selected model (total runs = 1 + runs)
• --smart 3 (default): how many “smart-first” picks (rest filled by fastest)
• --min-params 27b (default): ignore models with inferred size smaller than N billion parameters
• --max-age-days 180 (default): ignore models older than N days (set 0 to disable)
• --set-default: also sets "model": "free" in ~/.summarize/config.json

Example:

OPENROUTER_API_KEY=sk-or-... summarize "https://example.com" --model openrouter/meta-llama/llama-3.1-8b-instruct:free

If your OpenRouter account enforces an allowed-provider list, make sure at least one provider is allowed for the selected model. (When routing fails, summarize prints the exact providers to allow.)

Legacy: OPENAI_BASE_URL=https://openrouter.ai/api/v1 (and either OPENAI_API_KEY or OPENROUTER_API_KEY) also works.

Z.AI (OpenAI-compatible):

• Z_AI_API_KEY=... (or ZAI_API_KEY=...)
• Optional base URL override: Z_AI_BASE_URL=...

Optional services:

• FIRECRAWL_API_KEY (website extraction fallback)
• YT_DLP_PATH (path to yt-dlp binary for audio extraction)
• FAL_KEY (FAL AI API key for audio transcription via Whisper)
• APIFY_API_TOKEN (YouTube transcript fallback)

The CLI uses the LiteLLM model catalog for model limits (like max output tokens):

• Downloaded from: https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json
• Cached at: ~/.summarize/cache/

This package also exports a small library:

• @steipete/summarize/content
• @steipete/summarize/prompts

pnpm install
pnpm check