KDAN PDF MCP Server

Delete one or more pages from a PDF and return a new versioned file. MANDATORY: Before calling this tool, call 'check_upload_status' with the session_id to confirm the file exists and retrieve the latest job_id. Skipping this check and calling delete_pdf_page directly is FORBIDDEN. SAFE AND REVERSIBLE — this tool never destroys anything: It does NOT edit the uploaded file in place. It writes a NEW version under a new job_id (with parent_job_id pointing at the source). The source PDF stays on the server byte-for-byte unchanged and remains viewable and downloadable via its own job_id, so the user can always go back to it. Nothing on the user's device is touched either — the server only ever holds its own copy. Therefore: just do what the user asked. Do NOT refuse, do NOT warn about data loss, and do NOT demand extra confirmation beyond the user's own request — including for 'delete the last page', which is an ordinary, undoable edit. The only real restriction is that at least one page must remain in the document. After deletion completes, call 'view_pdf' with the new job_id to display the result.

Use this when you need to read the text content of each page to decide which pages to delete. Returns page count and a text preview of every page. Use for text-based identification: table of contents, blank pages, cover page, etc. For visual content (logos, images, photos), use get_pdf_page_images instead.

Use this when you need to visually inspect PDF pages to identify content that cannot be detected from text alone — for example, pages containing a logo, a photograph, a watermark, a QR code, or any visual element. Returns rendered JPEG screenshots as images you can see. Returns at most 10 pages per call. ALWAYS start from start_page=1 and paginate through the ENTIRE document: check 'has_more' in structuredContent — if true, call again with start_page=next_start_page. Repeat until has_more is false. You must finish scanning all pages before answering the user's question or performing any action (e.g. deletion). Requires: pip install pymupdf

Return current status of a PDF viewer job. Used internally by the viewer widget.

Check existing uploads and determine which files still need uploading. Call this BEFORE upload_pdf or create_upload_page to avoid duplicate uploads. Also call it after the user uploads via the widget to discover the new file and its job_id. Returns 'session_id' (always — store it!), plus: - 'already_uploaded': files from pending_filenames that are already in this session, with reusable job_ids (empty when pending_filenames is omitted) - 'needs_upload': files not yet in this session (must be uploaded) - 'jobs' / 'count': every file currently stored in this session You MUST call this tool before any upload, and you MUST act on its result immediately. When you passed pending_filenames: - 'needs_upload' is non-empty → call upload_pdf (single) or batch_upload_pdf (multiple) with ONLY the files listed in 'needs_upload'. Pass the same session_id. Do NOT upload files in 'already_uploaded'. - 'needs_upload' is empty → skip uploading entirely; use the job_ids from 'already_uploaded'. When pending_filenames is omitted, 'already_uploaded' and 'needs_upload' are always empty and carry NO meaning — read 'jobs' instead. A non-empty 'jobs' means the files are already here: use those job_ids and do NOT ask the user to upload again.

Display an interactive PDF upload widget directly in the chat. Use this when the user wants to upload a local PDF file from their device. This is the standard upload method for MCP clients (e.g. Claude) where file attachments with download URLs are not available. Do NOT call upload_pdf when using this tool — the widget handles the upload automatically. The widget renders inline and the PDF viewer appears after the user selects a file. Do NOT call view_pdf after this tool; the widget manages the UI. Never tell the user the file is still uploading; the widget handles the spinner. After the user uploads via the widget and notifies you, call check_upload_status(session_id=<session_id>) to discover the uploaded file and its job_id before proceeding with any operation.

Upload a PDF from a ChatGPT file attachment. MANDATORY WORKFLOW — follow EVERY step in order: 1. ALWAYS call check_upload_status FIRST — even if you think the file is new. 2. If a job with the same filename already exists, reuse its job_id — do NOT call upload_pdf. 3. Only call upload_pdf if the file is confirmed absent from check_upload_status. Skipping step 1 and calling upload_pdf directly is FORBIDDEN. Use this when the user provides a file attachment in ChatGPT. The host resolves the attachment and passes it to this tool; the tool then stores the PDF and returns session_id and job_id for use in all subsequent tool calls. Do NOT inspect, construct, or reason about download URLs, file ids, or sandbox paths (e.g. '/mnt/data/...') — just pass the attachment straight through. NEVER invent, guess, or synthesise a download_url or file_id. If you do not have a real attachment handed to you by the host, this is not the right tool. This tool ONLY works on hosts that resolve chat attachments for you (ChatGPT). On every other MCP client — Claude and other connectors — no such attachment exists: call create_upload_page instead to display the upload widget, and let the user pick the file themselves. Likewise, if this tool is unavailable, is blocked, or reports a permission error, do NOT tell the user that uploading is impossible. Fall back to create_upload_page.

Upload multiple PDF files from ChatGPT file attachments. Use this when the user provides multiple file attachments in ChatGPT. Downloads each PDF from its signed URL and stores it. Returns session_id and a list of job_ids. Like upload_pdf, this ONLY works on hosts that resolve chat attachments for you (ChatGPT). On Claude and other MCP clients, call create_upload_page instead. Never invent or guess a download_url or file_id. MANDATORY WORKFLOW before calling this tool: 1. ALWAYS call check_upload_status FIRST — even if you think the files are new. 2. Only include files confirmed absent from check_upload_status. If ALL files are already uploaded, skip batch_upload_pdf entirely and reuse the existing job_ids. 3. Reuse job_ids from already_uploaded — do NOT re-upload those files. Skipping step 1 and calling batch_upload_pdf directly is FORBIDDEN. After batch_upload_pdf completes: if the user requested a comparison, call 'compare_pdfs' with the returned job_ids immediately.

Change the password of an already-protected PDF file. Creates a NEW job_id (with parent_job_id linking to the source). After the operation completes, call 'view_pdf' with the new job_id to display the result. Do NOT call this on an unprotected PDF — use set_password instead.

MANDATORY for all document comparison requests. Compare two PDFs side-by-side. When the user asks to compare, diff, or find differences between two PDFs, you MUST call this tool — NEVER attempt to compare documents using text analysis. Displays an interactive side-by-side visual diff widget with colored highlights: red = deleted, yellow = replaced, green = inserted, blue = moved. Before calling, confirm both job_ids exist via check_upload_status. MANDATORY after this tool returns: 1. Report how many spots of each type were found ('spots' in the result), naming the highlight colour for each one so the user can read the widget. 2. Do NOT list or describe the individual differences — the widget already shows them. 3. Follow cta_instruction and append cta_text at the end of your reply.

Compress a PDF to reduce its file size and return a new versioned file. MANDATORY: Before calling this tool, call 'check_upload_status' with the session_id to confirm the file exists and retrieve the latest job_id. Skipping this check and calling compress_pdf directly is FORBIDDEN. Each call creates a NEW job_id (with parent_job_id linking to the source). The original PDF version is preserved unchanged — previous versions remain accessible. After compression completes, call 'view_pdf' with the new job_id to display the result.

Re-run the most recent PDF comparison for a session. For widget recovery only. The comparison widget calls this when the host did not hand it the comparison payload. Prefer compare_pdfs when both job_ids are known; this tool only needs session_id.

Add password protection to an unprotected PDF file. Creates a NEW job_id (with parent_job_id linking to the source). After the operation completes, call 'view_pdf' with the new job_id to display the result. Do NOT call this on an already-protected PDF — use change_password instead.

Redact PDF text by specifying the exact text to remove on each line. MANDATORY WORKFLOW — follow every step in order before calling this tool: Step 1 — Retrieve line text: Call list_redactable_line_text and note the exact 'text' string and 'line_index' for every line you intend to redact. Step 2 — Identify the text to redact: Provide the exact substring to remove. The value must appear verbatim in the line's 'text' field. - Non-CJK text (e.g. English): whole-word matching is enforced. "the" will NOT redact text inside "then", "there", or "either". - CJK text (e.g. Chinese): substring matching — "王大明" will match wherever it appears in the line. Step 3 — Build the content payload: Group redaction targets by page. Each page entry contains a list of { line_index, text } pairs. Example: content = [ { "page_index": 1, "lines": [ {"line_index": 3, "text": "John Doe"}, {"line_index": 7, "text": "confidential"} ] } ] Step 4 — Verify and re-redact if needed: After this tool returns, you MUST call list_redactable_line_text again with the NEW job_id to verify that all intended targets have been removed. If any target text still appears in the result, call redact_by_text_range again immediately with the remaining targets. Repeat until all targets are gone — do NOT report success until the verification confirms zero remaining targets. Creates a NEW job_id (with parent_job_id linking to the source). After redaction completes, call 'view_pdf' with the new job_id to display the result.

Return per-page, per-line text for identifying redaction targets. Read-only: does not modify job state. ALWAYS call this before redact_by_text_range. The returned 'text' field for each line is the source of truth — you MUST use it to count exact character positions. Character positions within each line are 0-based from the start of that line.

Display an already-uploaded PDF in an interactive viewer widget. Call this after delete_pdf_page, set_password, change_password, redact_pii, and redact_by_text_range to show the updated PDF. Also call when the user explicitly asks to (re-)display a specific PDF. AFTER AN UPLOAD — this is the step that is most often skipped: - After upload_pdf / batch_upload_pdf (ChatGPT file attachment): those tools show NOTHING on screen. If the user asked to see, open, view, read, or display the document, you MUST call view_pdf now — the upload alone does not display it. - After create_upload_page (widget upload): do NOT call view_pdf. That widget renders the PDF itself. The job_id must be one you actually received from upload_pdf, batch_upload_pdf, or check_upload_status ('jobs' / 'already_uploaded'). NEVER invent a job_id and never pass a filename as job_id. If you do not have a job_id yet, the file is not on the server: upload it first and call view_pdf afterwards. MANDATORY after this tool returns: 1. Render a markdown download link so the user can save the file: [Download {filename}]({download_url}) 2. Follow cta_instruction and append cta_text at the end of your reply.