Load a URL in a headless browser (hyperbrowser, browserless, steel) and return the page as markdown.

Generate or edit an image via a logged-in web session. For a new image pass prompt; in local stdio mode, edit an existing image by passing its absolute path in file and describing the changes. To continue editing a ChatGPT-generated image in the same chat, pass the returned session back. Local stdio and the local hosted bridge save the result and return its path; use path to choose the local destination. Direct hosted HTTP returns inline artifacts without writing server paths.

Deep research over multiple sources (parallel, exa, tavily, tinyfish, linkup, you, gemini_web, grok_web, chatgpt_web); may take minutes. gemini_web and chatgpt_web first return a research PLAN plus a session: pass that session with query "start" to run it, or with a revised request to replace the plan. gemini returns the finished report in the same call; chatgpt runs ~5-30 min after "start" — call again with the returned session to fetch the report (model is ignored there). Knobs: topic, recency, domains, depth ("deep" is slower/pricier). Provider-specific extras → call usage(provider=…).

Read a URL and return its main content as clean markdown; auto-escalates to a headless browser when the plain read comes back empty. Provider-specific extras via mode (crawl, structured extract, screenshot…) → call usage(provider=…) for exact params & example calls.

Web search across quota-aware providers. API providers return ranked title/url/snippet results; web providers (gemini_web, grok_web, chatgpt_web) return a synthesized answer with sources and a session token for follow-up turns. For chatgpt_web this is a full chat turn with web search on by default — mode="chat" answers from the model alone. In local stdio mode, attach files with file to ask about them. Niche knobs: topic, recency, domains. Provider-specific extras (scholar/patents/places, structured extract…) → call usage(provider=…) for exact params & example calls.

Show remaining free quota per account and provider for the current period, plus each account's assigned proxy. For logged-in web sessions (grok/gemini/chatgpt) also includes live per-tier limits and the selectable model/mode catalog with thinking levels — a locked mode (e.g. grok Expert/Heavy on a lapsed sub) reads 0/0. Pass provider for that backend's full capability sheet — its niches, escape-hatch mode values, and ready-to-copy example calls.