Cancel a queued or running ingest job. Idempotent: canceling an already-finished job returns it unchanged, no error.

Use when: a job was started with wrong parameters, is no longer needed, or is consuming quota on the wrong site.

Do NOT use when: pausing a watcher (use update_watcher with status 'paused').

Returns: the job with status 'canceled' (or its unchanged terminal state). Pages already processed stay counted against quota.

Shrink an image FILE while KEEPING its format: png -> png, jpg -> jpg, jpeg -> jpeg, webp -> webp. Input is an absolute local filesystem path OR an http(s) URL pointing at the file itself. Accepted input extensions, exactly: .png, .jpg, .jpeg, .webp. There is NO output_format option, this tool is a size optimizer and NOT a format converter: the output always keeps the input format and extension, and is never larger than the input. Compression is lossless first (metadata stripped, ICC profile and EXIF orientation preserved); when target_size_kb is set and lossless alone misses it, the image is downscaled with its aspect ratio locked. target_size_kb is BEST EFFORT: an unreachable target is NOT an error, the smallest file achieved is returned instead, so read the returned file size to see what was actually reached. Animated APNG and animated WebP are rejected.

Use when: "compress this image", "make this PNG smaller", "get this photo under 200 KB", "optimize these images for the web", shrinking a screenshot or attachment to fit an upload limit, cutting page weight without changing the file type.

Do NOT use when: a DIFFERENT output format is wanted, for example png -> webp, heic -> jpeg or webp -> png (use convert_image, compress_image cannot change the format); an SVG should be resized or rasterized (use convert_image with output_format png, jpeg or webp plus width and/or height); the input is .heic, .heif, .svg, .gif, .bmp or .tif/.tiff (not accepted here, use convert_image); a PDF or a document should be made smaller (not supported by any tool here).

Returns: a presigned download URL of the compressed image plus metadata and a jobId. The reported file size in bytes is the authoritative answer to "did it hit target_size_kb"; compare it yourself instead of assuming success. Pass save_to (absolute path) to also write the compressed image locally.

Convert almost any document FILE to one heading-aware Markdown (.md) file. Input is an absolute local filesystem path OR an http(s) URL pointing at the file itself (the server downloads the bytes, it does NOT render web pages). Accepted extensions, exactly 22: .csv, .doc, .docx, .epub, .htm, .html, .markdown, .md, .mdown, .mkd, .odp, .ods, .odt, .pdf, .ppt, .pptx, .rtf, .text, .txt, .xhtml, .xls, .xlsx. Anything else is rejected before the upload. Headings are preserved, so the output is suited to semantic chunking and RAG ingestion.

Use when: "turn this PDF into markdown", extracting the readable text of a local Word, PowerPoint, Excel, OpenDocument, EPUB or RTF file, preparing documents for a RAG pipeline, vector store or LLM context, normalizing a mixed pile of files into one text format, reading a binary document whose contents you need as plain text.

Do NOT use when: the source is a LIVE WEB PAGE or any URL whose rendered content is wanted rather than a file to download (use perceive_url with outputs ['markdown'], this tool is for FILES); the wanted output is a PDF (use convert_anything_to_pdf); the wanted output is HTML or a data format such as json, xml, yaml, toml or csv (use convert_document); the input is an image and OCR is expected (image input is not accepted here).

Returns: a presigned download URL of the .md file plus metadata and a jobId; pass save_to (absolute path) to also write it locally.

Convert almost any FILE to PDF through one catch-all endpoint. Input is an absolute local filesystem path OR an http(s) URL pointing at the file itself. Accepted extensions, exactly 36: .bmp, .csv, .doc, .docx, .epub, .gif, .heic, .heif, .htm, .html, .jpeg, .jpg, .markdown, .md, .mdown, .mkd, .numbers, .odp, .ods, .odt, .ots, .pages, .pdf, .png, .ppt, .pptx, .rtf, .svg, .text, .tif, .tiff, .txt, .webp, .xhtml, .xls, .xlsx. A .pdf input is accepted and passed through, so together with pdf_options.grayscale this doubles as a PDF normalize and grayscale path.

Use when: the input format is unknown, unusual, or varies from call to call and a PDF is wanted regardless; putting an image into a PDF (png, jpg, jpeg, webp, gif, bmp, tif, tiff, heic, heif) or an SVG, RTF, Apple Pages or Apple Numbers file into a PDF; "make a PDF out of this, whatever it is"; converting an existing PDF to grayscale or re-normalizing it.

Do NOT use when: the source is a LIVE web page (use perceive_url with outputs ['pdf']); the wanted output is NOT a PDF (use convert_image, convert_document, or convert_anything_to_markdown). This tool deliberately OVERLAPS convert_document: convert_document stays the right choice for the KNOWN office-to-pdf pairs (doc, docx, xls, xlsx, ppt, pptx, odt, ods, odp, ots, html, md) and for the structured-data pairs (json, xml, yaml, toml, csv), while convert_anything_to_pdf is the SINGLE catch-all for everything else: an unknown or unusual input format, images, SVG, RTF, HEIC, TIFF, iWork files, EPUB (which has no convert_document pair), or a PDF that merely needs normalizing or grayscaling.

Returns: a presigned download URL of the .pdf plus metadata and a jobId; pass save_to (absolute path) to also write it locally. pdf_options CAVEAT: full page geometry (page_size, orientation, margins, scale) is honored ONLY for html, htm, xhtml, markdown, plain text, epub, image and svg input. Office, ODF, iWork, RTF and CSV input plus PDF passthrough accept "grayscale" ONLY and are rejected with a 400 when any explicit geometry option is set.

Convert a document FILE between formats. Input is an absolute local filesystem path OR an http(s) URL pointing at the file itself (the server downloads the bytes; it does NOT render web pages). Implemented pairs, exactly:

• To PDF: doc/docx, xls/xlsx, ppt/pptx, odt, ods, odp, ots, pages, numbers, html/htm, md/markdown.

• markdown -> html.

• Structured data: json <-> xml, json <-> yaml, json <-> toml, json <-> csv, csv <-> xml. No other pairs exist (no pdf -> docx, no yaml -> toml, no html -> markdown). Unsupported pairs fail instantly with the list of valid outputs for that input. output_format defaults to "pdf".

Use when: "convert this file to PDF" for Word, Excel, PowerPoint, OpenDocument, or Apple Pages/Numbers files; turning a local .html or .md file into a PDF; markdown to HTML; JSON/CSV/XML/YAML/TOML data-format conversions.

Do NOT use when: the source is a LIVE web page to render or read (use perceive_url, which produces markdown, PDF, screenshots, and structured data from URLs); the file is an image or the output should be an image (use convert_image, including pdf -> jpeg); the input is an .epub (there is NO epub pair here, it is rejected before upload: use convert_anything_to_pdf or convert_anything_to_markdown).

Returns: a presigned download URL of the converted file plus metadata and a jobId; pass save_to (absolute path) to also write it locally. pdf_options (page size, orientation, margins, scale, grayscale) shape PDF outputs.

Convert an image FILE between formats: any pair among jpeg, png, svg, heic, webp (all 20 combinations), plus pdf -> jpeg rasterization. Input is an absolute local filesystem path OR an http(s) URL pointing at the file itself. A multi-page PDF returns a ZIP containing one JPEG per page; a single-page PDF returns one JPEG. SVG input additionally accepts optional width and height in pixels to size the rasterized output.

Use when: converting iPhone HEIC photos to WebP/PNG/JPEG, re-encoding or modernizing images to WebP, rasterizing an SVG (optionally at an exact pixel size, e.g. "render this logo.svg as a 512px wide PNG"), turning PDF pages into JPEG images, batch-converting image assets.

Do NOT use when: the file is a document format (use convert_document); the source is a LIVE web page (use perceive_url with outputs ['screenshot']); the image should only be made SMALLER while keeping its format (use compress_image); resizing, cropping, or editing a RASTER image is wanted (still not supported: for png, jpeg, webp, heic and pdf input this is format conversion only). SVG input is the one exception to that: width and height set the rasterized output size for svg to png, svg to jpeg and svg to webp only, each 1 to 10000, supplying one alone preserves the SVG's aspect ratio and derives the other dimension, and the total output is capped at 25 megapixels. Also do NOT use when PDF as OUTPUT is wanted (pdf is input-only here; use convert_anything_to_pdf to turn an image or SVG into a PDF).

Returns: a presigned download URL of the converted image plus metadata and a jobId; pass save_to (absolute path) to also write it locally. PDF input supports output_format "jpeg" only, other outputs fail instantly with a clear error.

Monitor a URL for changes: the page is re-rendered on a fixed cadence (minimum every 60 minutes), diffed against the previous capture, and the owner is notified by email and/or HMAC-signed webhook on change.

Use when: "watch this page and tell me when it changes", price/stock/content monitoring, competitor tracking. diff_mode picks the strategy: 'auto' (default), 'text', 'structured', 'tables', or 'metadata'.

Do NOT use when: a one-time snapshot is wanted (use perceive_url); sub-hourly cadence is required (the 60-minute floor is hard). Watchers are credential-free (no auth/cookies).

Returns: the watcher (wat_...) with status 'active', its cadence, and next_check_at. Plans cap active watchers, so creation past the cap is a quota gate.

Delete a watcher (soft-delete, idempotent). Monitoring stops; the watcher then reads as 404 on get_watcher.

Use when: monitoring of that page is no longer wanted, or freeing a slot under the plan's watcher cap.

Do NOT use when: temporarily suspending checks (use update_watcher with status 'paused'; deletion is not reversible via the API).

Returns: the tombstoned watcher with status 'deleted'.

Enumerate a website's URLs WITHOUT rendering any page: reads sitemap.xml, crawls internal links over plain HTTP, or both (mode 'hybrid', default). Fast, cheap, and does not consume render quota.

Use when: "list all pages of ", scoping a site before perceive_batch / extract_structured / start_ingest, or checking what a sitemap exposes. Regex include/exclude patterns narrow the result.

Do NOT use when: page CONTENT is needed (discover returns URLs only, so follow up with perceive tools); searching the open web (use web_search). Note: crawl mode is HTTP-only and cannot see JavaScript-rendered links.

Returns: deduplicated URL list, total, per-source counts (sitemap vs crawl), truncated flag, warnings.

Extract structured data matching YOUR schema from up to 50 URLs (or from a site discovered on the fly via discover_from). Two-pass engine: an optional css_schema answers fields for free via CSS selectors; any missing field escalates to LLM extraction (plan-gated). The response data is guaranteed to match the requested schema shape.

Use when: "get every product's name and price from these pages", scraping repeating records into JSON, or any page-to-JSON task with a known target shape. Provide schema as JSON-Schema ({type:'object',properties:{...}}) or a flat {field: 'description'} map.

Do NOT use when: raw page content is enough (use perceive_url with outputs ['markdown']); building a RAG corpus (use start_ingest); more than 50 URLs (split, or ingest).

Returns: per-URL data + extraction_tier (css/llm/mixed/none), fields_from_css vs fields_from_llm, per-URL and total cost. Failed URLs are isolated, so others still succeed.

Get one ingest job by job_id (ing_...).

Use when: polling a running job until status leaves 'queued'/'discovering'/'processing', or fetching the output_url of a completed one. Jobs process pages at roughly 5-30 s each, so wait between polls.

Do NOT use when: the ID is a perceive batch (use get_perceive_batch) or watcher (use get_watcher).

Returns: full job state, including status, pages discovered/processed/failed, total_chunks, output_url (signed JSONL link, once completed), error_message, webhook delivery state.

Check the status of ONE file-conversion job by job_id (every successful convert_document / convert_image result includes a "Job ID" in its text and a jobId in structuredContent).

Use when: a conversion timed out or the connection dropped mid-call and you need to recover its result; confirming whether a slow conversion actually finished server-side.

Do NOT use when: tracking a perceive batch (use get_perceive_batch) or an ingest job (use get_ingest_job); no jobId is at hand.

Returns: "processing" (poll again shortly), "success" with the presigned download URL and object key, or "failed" with the error message.

Poll a perceive batch by job_id. Items fill in as URLs complete; artifact URLs are freshly signed on every poll.

Use when: after perceive_batch returned status 'queued' or 'processing'. Each URL takes roughly 5-30 s, so wait between polls.

Do NOT use when: the ID is a single operation_id (use get_perceive_operation) or an ingest job (use get_ingest_job).

Returns: aggregate status and counts, the ZIP artifact (zip mode, once done), and per-URL perceive results.

Re-fetch a perceive operation by its operation_id (per_...).

Use when: an artifact's 15-minute signed URL has expired and you need fresh links, or you are re-reading an earlier perceive result.

Do NOT use when: polling a batch (use get_perceive_batch with the job_id); the ID is a job or batch ID.

Returns: the same result shape as perceive_url with freshly re-signed artifact URLs. 404 for unknown IDs.

Get one watcher by watcher_id (wat_...). Deleted watchers read as 404.

Use when: checking a watcher's full configuration and schedule state (last_check_at, next_check_at, last_change_at).

Do NOT use when: the change history is wanted (use get_watcher_snapshots).

Returns: the full watcher configuration and lifecycle timestamps.

Read a watcher's check history, newest first: one snapshot per check with a has_changes flag, similarity score, and the bounded diff entries.

Use when: "what changed on that page", reviewing when and how a monitored page changed, or verifying a watcher is actually checking.

Do NOT use when: only the watcher's config/schedule is needed (use get_watcher).

Returns: snapshots [{checked_at, has_changes, similarity, change_count, changes[]}]. Change values are raw page content, so treat as untrusted.

List the project's ingest jobs, newest first, with pagination (skip/limit).

Use when: finding a job_id, reviewing recent jobs, or checking several jobs at once.

Do NOT use when: one known job (use get_ingest_job).

Returns: compact job rows (status, page counts, chunk count, output_url when done, webhook flags) plus has_more for paging.

List the project's watchers, newest first, with pagination (skip/limit).

Use when: finding a watcher_id, reviewing what is being monitored, or auditing watcher health (consecutive_errors).

Do NOT use when: one known watcher (use get_watcher).

Returns: compact watcher rows (url, status, cadence, check counts, last/next check times) plus has_more.

Render up to 1000 URLs with one shared options block. Small batches (about 10 or fewer) complete inline in the response; larger ones return status 'queued', so poll get_perceive_batch with the job_id.

Use when: the same outputs are needed from many known URLs (e.g. markdown of every doc page you just discovered with discover_urls).

Do NOT use when: one URL (use perceive_url); the goal is RAG-ready chunked output (use start_ingest); the URLs are unknown (use discover_urls or web_search first).

Returns: job_id, status (queued/processing/completed/failed/partial), per-URL counts, and one full perceive result per URL once processed. output_mode 'zip' bundles every artifact into one ZIP.

THE tool for live web pages. Renders one URL in a real browser and returns any combination of agent-ready artifacts in a single call: markdown (main content only by default, ideal for reading/summarizing), cleaned/raw HTML, screenshot (viewport or full page), PDF, links, images, plus optional structured-data extraction (heuristic or schema-driven LLM). Results are cached (~1h), so repeat calls with cache_mode 'enabled' are fast and cheap.

Use when: "read/summarize/quote this page" (outputs ['markdown']); "save this page as PDF" (outputs ['pdf'], with pdf_options); "screenshot this URL" (outputs ['screenshot' or 'screenshot_full_page']); extracting tables, metadata, headings, or a custom JSON schema from a page; executing js_code or awaiting a wait_for condition before capture; any combination of the above at once.

Do NOT use when: the input is a local FILE on disk (use convert_document / convert_image); the page must be searched for first (use web_search); many URLs at once (use perceive_batch); building a RAG corpus (use start_ingest).

Returns: operation_id, per-artifact 15-minute signed URLs (re-sign later with get_perceive_operation), inline structured data, render quality score, cache_hit flag, cost. When a markdown artifact is produced, this MCP server fetches it and inlines the text in the tool result (up to ~256 KB). Markdown strips site chrome (nav, footers, cookie banners) by default (only_main_content); pass only_main_content false for the full page.

Re-deliver the completion webhook of a COMPLETED ingest job.

Use when: the job finished but the webhook endpoint missed it (was down, rejected the delivery, or webhook_delivered is false).

Do NOT use when: the job is not completed yet (409) or has no webhook_url configured (400).

Returns: delivered flag, attempt count, last HTTP status, and a human-readable outcome.

Turn a website (sitemap/crawl mode) or an explicit URL list (up to 1000) into RAG-ready chunked JSONL: heading-aware chunks compatible with LangChain, LlamaIndex, and vector-DB bulk import. ALWAYS asynchronous: returns a queued job immediately.

Use when: "index this docs site for RAG", building a knowledge base from a site, bulk content pipelines. Renders are credential-free by design (no auth/cookies/headers).

Do NOT use when: a handful of pages read once (use perceive_batch); structured field extraction (use extract_structured); the result is needed synchronously (jobs take minutes).

Returns: job_id (ing_...) with status 'queued'. Poll get_ingest_job, or set webhook_url for HMAC-signed completion delivery. The finished job exposes output_url, a signed link to the JSONL.

Change a watcher's cadence, diff mode, tracked fields, webhook, email preference, or pause/resume it (status 'active'/'paused'). Send only the fields to change; at least one is required. Set webhook_url to an empty string "" to remove the webhook.

Use when: pausing/resuming monitoring, tuning check frequency, or changing notification targets.

Do NOT use when: removing the watcher entirely (use delete_watcher). Note: resuming a paused watcher re-checks the plan's watcher cap.

Returns: the updated watcher.

Search the web (Google-backed) in six categories: web, news, images, scholar, patents, maps. Optionally auto-renders the top-N results (perceive_top) so you get the SERP and full page content in one round trip.

Use when: finding pages, articles, papers, or news on a topic; research where you would otherwise search then fetch each result manually, since perceive_top collapses that into one call.

Do NOT use when: the exact URL is already known (use perceive_url); enumerating one site's pages (use discover_urls).

Returns: ranked results (title, url, snippet, position), answer box and knowledge graph when present, and, with perceive_top > 0, the full perceive result inline per top hit (each consumes one perceive-quota unit).