summarize

Why it's ranked here

The breadth is real and documented down to the fallback order. It can drive nine installed coding assistants as model backends, including Claude, Codex, Gemini and Copilot, plus direct keys for providers such as OpenAI, Anthropic, Google and xAI, and a local Ollama option. Auto mode skips models with no key, retries the next candidate on any error, and prints the plain extracted text if nothing is usable. Repeat runs are cheap: transcripts, extractions and finished summaries are cached, keyed on the actual content sent. MIT licensed, installable from npm or Homebrew, with a separate core library.

What's good

Transcripts come first: YouTube goes through two caption sources before downloading audio, and only then walks a stated order of transcription services. The summary cache is keyed on a hash of the content and prompt, so the same article under two addresses hits once. The background service binds to the loopback address, requires a token on every route except health, and applies an auth rate limiter. When a coding assistant is the backend, Codex runs in a throwaway home directory carrying only auth, and pi runs with tools, extensions and context files switched off.

Tradeoffs

Requires Node 24 or newer for the npm install, which rules out many older machines unless you use the Homebrew or macOS builds. Media depth leans on outside binaries: slide capture needs yt-dlp, OCR needs tesseract. To summarize a local image or binary file through Claude Code, it passes the flag that skips permission prompts, and Gemini gets its yolo mode. The docs say the Antigravity no-tools guidance is prompt steering, not a hard boundary. The built-in summary prompt forbids quotation marks and sponsor mentions, which you replace wholesale if you supply your own.

How to use it well

Best for someone who already pays for a coding assistant subscription and wants long videos, podcasts and articles condensed without setting up new API keys: point the auto fallback at that assistant. Use the extract mode when you want clean text to feed your own pipeline, and the stable JSON output for scripts. Install the background service only if you need OCR, native media tools or Firefox media support; the browser sidebar works on its own. It is a reader, not a research agent: it summarizes one source at a time.

Observed

License

MIT (LICENSE)

Language

TypeScript, ES modules (package.json)

Runtime requirement

Node.js 24 or newer (package.json engines)

Install surface

npm global package, Homebrew, npx one-shot, macOS archives (README.md)

Interfaces

CLI, Chrome side panel and Firefox sidebar, local HTTP daemon, core library package (README.md)

Model backends

Direct provider API keys, OpenAI-compatible endpoints, Ollama, OpenRouter, and installed coding CLIs (docs/cli.md, docs/llm.md)

Daemon binding

Loopback address, default port 8787, token-authenticated (docs/chrome-extension.md, src/daemon/server.ts)

Cache

Single SQLite file with size cap and TTL eviction (docs/cache.md)

Test runner

Vitest with coverage in the check script (package.json)