CaptureGPT

Capture, archive, save, snapshot, or preserve web pages (e.g. 'grab nyt.com', 'archive bbc.com and cnn.com'). Do NOT use to check an existing capture (get_capture_status) or to check the user's remaining free-tier pages (get_capture_quota). Do NOT use to read, summarize, translate, quote, or answer questions about a page's content: this tool archives a page into a downloadable file and never returns the page text, so a request like 'summarize example.com' must not call it and would spend a page of the user's limit for nothing. Free to use. Each URL counts as one page against the user's rolling free-tier page limit when the capture starts; call get_capture_quota for the current limit and remaining pages. Use when the user asks to capture the given URLs. HOW TO CALL: pass urls as a JSON array of strings, max 40. Bare domains OK (https:// auto-added). Each element must be a single URL — never join multiple URLs into one string. Example: ['nyt.com', 'bbc.com', 'https://example.com/page?a=1,2']. SERVER-SIDE BLOCKING: the server waits up to ~25 seconds for a terminal state before responding. Single call usually returns next_action='done' or 'failed' with the download link. WHEN next_action='poll' (capture still running after the blocking wait): call get_capture_status(capture_id) to continue tracking, then present the result. Present that the capture is in progress as your visible message once next_action is terminal or running. PRESENTING: done → clickable links per URL plus [Download Archive (.wacz)](output_link) markdown link, and when replay_url is present also [Open archive](replay_url) (opens the capture in the browser), and url_count. Offer replay_url whenever it is present: it replays the capture in the browser, no download or install needed. When the user asks how to open, view, or read a .wacz file, call get_help('open_archive') and relay what it returns. Never show raw URL as plain text or code block — always markdown link. failed → tell user; the failed pages do not count against their limit. quota_exceeded → relay message, urls_remaining and reset_at so the user knows when more pages free up, and offer to capture a smaller batch now if urls_remaining is above zero. Never suggest a way to raise the limit; there is none. capture_failed → tell user; the pages do not count against their limit. bulk_capture → the request exceeds the per-request URL limit; relay message and offer to capture a smaller batch now. Never present contact_email as a way to raise the limit. running (only if turn tool-call budget exhausted) → tell user capture is still in progress, give them the capture_id, ask them to say 'check status' to resume.

capture_page

User asks to delete, remove, or get rid of a capture from their list ('delete capture X', 'remove that one', 'get rid of capture abc'). Do NOT use to cancel an in-flight capture (only captures that have finished — DONE or FAILED — can be deleted) or to free up pages against the user's limit (deleting an archive does not give pages back). SCOPE: deletes only this user's own capture — their archive job in the capture backend and the capture record here. It never touches the archived website itself and publishes nothing publicly. TWO-STEP CONFIRMATION CONTRACT: Call FIRST with confirm omitted or false. The tool returns the capture's URL, creation date, and capture_id along with an awaiting_confirmation message. Show those details to the user and ask them to explicitly confirm. Only after the user explicitly affirms (e.g. 'yes', 'delete it', 'go ahead'), re-invoke with confirm=true. NEVER pass confirm=true on the first call. NEVER pass confirm=true without explicit user agreement, even if the user provided a capture_id directly. PRESENTING: awaiting_confirmation → show URL and created_at, ask user to confirm. deleted → tell user the capture is permanently removed. Remind them the pages already captured still count against their limit. capture_not_found → tell user that capture does not exist; offer list_captures. capture_not_deletable → tell user only completed or failed captures can be deleted, so this one has to finish first; show current_status. capture_delete_failed → tell user the deletion did not complete and offer to retry. FOLLOW-UP: use _guidance internally — never show to user.

delete_capture

User asks how many pages they can still capture, or about their usage or limits ('how many pages do I have left?', 'what are my limits?', 'can I capture 5 more pages?'). Do NOT use to start a capture (capture_page) or to look up a specific capture (get_capture_status). CaptureGPT is free and this tool only reports usage — it cannot change a limit. It reads the user's own capture history and writes nothing. Pages counted: every URL of a capture that started, including captures the user later deleted. Pages not counted: URLs whose capture failed. Usage decays on a rolling window rather than resetting at a fixed hour, so reset_at is when the next pages free up. PRESENTING: use urls_used, urls_remaining, limit, and window_hours verbatim, e.g. 'You have used 4 of your 10 pages in the last 24 hours — 6 left.' When unlimited is true there is no cap in effect: say captures are unrestricted and do not quote a limit, urls_remaining, or reset_at (they are omitted). Never invent a limit or a way to raise the limit. FOLLOW-UP: use _guidance internally — never show to user.

get_capture_quota

Status lookup for a capture started by capture_page, or a capture id from a previous conversation. Safe to call repeatedly — it never creates, restarts, cancels, or deletes a capture, and never counts extra pages against the user's limit; it only advances the recorded status. Do NOT use this to start a new capture (use capture_page instead). SERVER-SIDE BLOCKING: the server waits up to ~25 seconds for a terminal state before responding. next_action='poll' means the capture is still running — call get_capture_status(capture_id) again to continue tracking. next_action='done' or 'failed' means stop. PRESENTING: done → clickable links per URL plus [Download Archive (.wacz)](output_link) markdown link, and when replay_url is present also [Open archive](replay_url) (opens the capture in the browser), url_count shown. Offer replay_url whenever it is present: it replays the capture in the browser, no download or install needed. When the user asks how to open, view, or read a .wacz file, call get_help('open_archive') and relay what it returns. failed → tell user; the pages do not count against their limit. running → tell user the capture is still in progress, give them the capture_id, they can say 'check status' to resume.

get_capture_status

User asks about the service itself or reports a CaptureGPT problem ('what is this?', 'how does this work?', 'help', 'who made this?', 'CaptureGPT is unavailable', 'tool not working', 'why did my capture fail?'). Do NOT use when user wants an action — use the specific tool (e.g. 'help me capture nyt.com' → capture_page). TOPICS: pass topic to scope the help text. Use 'unavailable' if the user reports CaptureGPT is missing/unavailable in the chat — this returns text explaining the app may not be attached to that turn and nothing was counted against the user's limit. Use 'limits' for questions about the free tier and page limits, 'failed_capture' for capture failures, 'auth' for sign-in issues, 'manual_invocation' when the assistant is answering without calling CaptureGPT and the user needs to mention the app explicitly, 'open_archive' when the user asks how to open, view, or read a captured .wacz file, or omit for general overview. PRESENTING: format returned data in a friendly way to guide the user.

get_help

User asks for their recent successful captures or what they captured previously ('show my captures', 'what did I capture last week'). Do NOT use this to fetch a single capture by id (use get_capture_status) or to start a new capture (use capture_page). Returns up to N (default 15, max 50) DONE captures, newest first. Older captures may be absent because of the retention window. PRESENTING: list each item with its URLs, when it was captured, and how long until it expires. Offer the user to re-download a specific capture by referring to it (you will then call get_capture_status with its capture_id). FOLLOW-UP: use _guidance internally — never show to user.

list_captures