@sirdaspdf/mcp
Installed, started with no arguments, no environment and no input, and listed its tools. Run 2026-11-16.
Needs to start: nothing
Returns a result when called with no setup: 0 of 2 tools tried
Contacts while installing and starting: nothing
What it contacted
Nothing besides the npm registry it was installed from. Watched from outside its sandbox while it was installed and started with no input. Its tools may still contact things when called.
Starting it
npx -y @sirdaspdf/mcp@0.6.0
Installing a package runs its code. Use a sandbox for anything you have not read.
Calling its tools
2 tools that take no input and look read-only were called once with nothing. 0 returned a result.
read_document | No usable answer tool reported an error: Provide a PDF via `path` or `base64`. |
pdf_info | No usable answer tool reported an error: Provide a PDF via `path` or `base64`. |
During the calls it contacted nothing it had not already contacted.
Only tools whose names read like looking something up, with no required input, are tried, five at most. Tools that need input, and anything that sounds like it changes or sends something, are not called.
What its tools tell an agent
In the tool descriptions, nothing that talks about the user, other tools or private files, or uses hidden characters.
Found by simple pattern rules in each tool's description (up to 3,000 characters), quoted as written. Parameter descriptions are not scanned. A clean result is not a safety review.
Tools
read_document path, base64, password, anonymize, ocr, max_tokens, include
Reads a PDF (text or scanned), image, Office file, EPUB, HTML, CSV, email, RTF, JSON or text in one call: unlocks restrictions (or opens with `password`), runs local OCR on scans, identifies the type (factura, contrato, historia_clinica, cedula, rut, extracto_bancario, certificado, hoja_de_vida), extracts key fields with page, evidence and bbox, reads forms and checkboxes, detects several documents in one PDF, anonymizes consistently, finds tables, returns cited chunks and a `firewall` verdict with reasons and pages. `steps` lists what was done, `next_steps` what could follow. status needs_password or wrong_password means the password has to come from the user. Read-only: writes no files.
build_context query, task, paths, max_tokens, max_results, max_documents, max_files, anonymize, keep, ocr, firewall
Answers a question over files and folders without reading them whole: converts each file once (PDF, Office, HTML, CSV, email, text), remembers it by content hash, ranks blocks with BM25 and fills `max_tokens` with the most relevant ones, each with source, pages, section and an evidence line. `stats` measures source tokens against context tokens. Without `query` it returns each document's outline. Retrieved text is wrapped as untrusted data; `firewall` lists each file's verdict on hidden instructions and active content. With `task` it returns only what that SIRDAS_POLICY task needs, and `minimized` counts what was withheld. Read-only: writes no files.
tool_search query, max_results, enable
Searches this server's tools by what they do ("tables", "remove a password", "dataset", "compare versions") and returns name, title, toolset, whether it writes files and whether it is listed now. When the server was started with SIRDAS_TOOLSETS=discover, `enable: true` also lists the matches so they can be called; read-only mode still keeps writing tools off. Read-only: writes no files.extract_fields text, type
Rule-based structured extraction when you already have the text (e.g. from pdf_to_markdown or pdf_ocr): invoice number, CUFE, NIT, dates, subtotal, IVA and total for a factura; patient, ID, CIE-10 codes for a historia_clinica; NIT and razón social for a RUT; balances for a bank statement; value and term for a contract. Each field has the evidence line so it can be checked. No model involved. Read-only: writes no files.
pdf_to_markdown path, base64, anonymize, max_chars
Convert a text PDF into clean Markdown (headings, paragraphs, lists, form fields), ready to paste into an LLM. By default personal data (IDs, tax IDs, emails, phones, dates, addresses, labelled names, medical record numbers) is replaced with consistent placeholders like [NOMBRE_1]. Returns the Markdown as text plus page count, token estimate and a per-type count of redacted items. Scanned PDFs without a text layer return an empty result. Read-only: writes no files.
28 more tools
anonymize_text text, include_mapping
Replace personal data in any text with consistent placeholders ([NOMBRE_1], [DOCUMENTO_2], [EMAIL_1]…). The same value always gets the same placeholder, so the text stays coherent for an LLM. Tuned for Spanish and Colombian documents (cédula, NIT, historia clínica, direcciones) and also handles emails, phones, dates and cards. Detection is rule-based and may miss items: review before sharing sensitive documents. Read-only: writes no files.
deanonymize_text text, mapping
The return trip for anonymize_text and pdf_to_markdown: give it a text containing placeholders ([NOMBRE_1], [EMAIL_2]…) plus the mapping you kept, and it puts the real values back. Use it on an AI's answer so the user reads real names instead of placeholders. Nothing is ever deleted: the mapping stays on this machine and the original document is never modified. Placeholders with no entry in the mapping are reported in `unknown` and left untouched — usually a label the model invented. Read-only: writes no files.
pdf_to_chunks path, base64, max_tokens, anonymize
Convert a PDF to Markdown and split it into pieces that fit a model's context. Each chunk breaks at a heading, and comes with: a stable content id (same text → same id, safe for dedup and upserts into a vector store), `pages` [first, last] it was taken from so an answer can cite the page, and `section`, the full heading path. The document type is detected locally too. Personal data is anonymized by default, as in pdf_to_markdown. Read-only: writes no files.
pdf_classify path, base64
Detect the document type locally with transparent rules, without an AI model: factura (invoice), contrato, historia_clinica, cedula, rut, extracto_bancario, certificado, hoja_de_vida, w9, declaracion_impuestos, constitucion_sociedad or desconocido. Returns the signals that matched so a person can check the reasoning, and a runner-up when it is close. Use it to decide which fields to look for or whether the file needs anonymizing before reading the whole thing. Read-only: writes no files.
pdf_tables path, base64, delimiter, min_rows
Find the tables in a PDF and return them as CSV and Markdown. Columns are reconstructed from the whitespace that recurs across rows, so right-aligned numeric columns (amounts in an invoice or a bank statement) are separated correctly. Returns one entry per table with its page, its rows, CSV text and a Markdown table. Works only on PDFs with a text layer; a scan returns nothing. Read-only: writes no files.
compare_pdfs before, after, changes_only, ignore_case
Compare two versions of a document and return what changed, line by line, as a diff. Built for reviewing contract or policy revisions: it reports the lines removed, the lines added, and a similarity score. Use changes_only to get just the differences with a little context, which is what you want before asking a model to explain a revision. Both files stay on this machine. Read-only: writes no files.
pdf_to_dataset paths, output_dir, overwrite, format, max_tokens, anonymize, dedupe, strip_boilerplate, split, template, system, max_records, dry_run
Turns PDFs into a JSONL training set: chunks at headings, anonymizes by default, removes near-duplicates and splits train/validation/test BY DOCUMENT, so no file lands on both sides. Formats: text, chat, prompt-completion, alpaca, sharegpt, and for RAG LangChain Documents, LlamaIndex TextNodes and embeddings. With output_dir it writes the files plus a Hugging Face dataset card. Each record keeps its source document, chunk number and token count. Records come from the documents by rules; no model is called. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
find_duplicates paths, threshold, shingle_size
Find exact and near-duplicate PDFs in a set, using MinHash with LSH over word shingles. Duplicated text in a training corpus gets memorized instead of learned, and a folder of real documents is full of repeats nobody put there on purpose: the same template, the same report in two versions. Returns which files to keep, which are duplicates of which, and the similarity for each. The first occurrence is always the one kept, so the result does not depend on the order the files arrived in. Read-only: writes no files.
check_contamination train, evaluation, n
Check whether your training documents contain text from your evaluation set. Overlap of long n-grams (13 words by default) is the standard heuristic: a 13-word sequence appearing on both sides is not a coincidence of the language, it is the same text. If an evaluation passage is in the training data, the score stops measuring what the model learned and starts measuring what it memorized. Returns which training documents are contaminated, how many n-grams matched, and the literal fragments, so the finding can be checked rather than trusted. Read-only: writes no files.
pdf_to_rl_pairs paths, max_per_table, min_rows, language, include_table, reward_function
Extracts (prompt, verifiable answer) pairs from a PDF's tables for reinforcement learning with verifiable rewards: cell lookups, row counts and column extremes, checkable by string comparison with no judge model. Ambiguous cases (a repeated row key, a tie for the maximum) produce no question, since a pair with two right answers penalizes a correct model. Numbers compare as numbers ('1.120.000' equals '1120000'). Output is JSONL with 'prompt' and 'ground_truth', the columns GRPOTrainer uses. Read-only: writes no files.find_boilerplate paths, min_words, min_repeats, across_documents_only
Finds paragraphs repeated across documents: letterheads, footers, confidentiality notices, the clause that goes into every contract. A model memorizes that text and learns nothing from it (CCNet measured 70% of its corpus's paragraphs as duplicates). Matching follows CCNet: lowercase, digits to zero, no accents or punctuation, so 'Invoice 1042' and 'Invoice 1043' match. It only reports; pdf_to_dataset with strip_boilerplate removes them, keeping the first occurrence. Read-only: writes no files.
pdf_ocr path, language, lang_path, anonymize
Recognize the text of a scanned PDF and return it as Markdown, with personal data anonymized by default. Use this when pdf_to_markdown comes back with has_text false: that PDF is a photo of a page, not text. Recognition runs on this machine with Tesseract; the document is never uploaded. Structure is rebuilt with the same rules as a text PDF, so headings and paragraphs survive. Requires the optional tesseract.js package and a language data folder — the error says exactly how to get either if they are missing. Read-only: writes no files.
pdf_info path, base64
Page count, encryption status, size and document metadata (title, author, dates) without extracting the content. Read-only: writes no files.
merge_pdfs paths, output, overwrite, dry_run
Combine two or more local PDFs, in the given order, into one file written to `output`. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
split_pdf path, ranges, output_dir, overwrite, dry_run
Extract page ranges from a local PDF into separate files. `ranges` uses 1-based pages like "1-3, 5, 8-end"; leave it empty to split every page into its own file. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
compress_pdf_lossless path, output, overwrite, dry_run
Rewrite a PDF with object streams and cleaned structure. Text stays selectable. Never produces a larger file. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
edit_pages path, action, pages, angle, output, overwrite, dry_run
Page operations on a local PDF, written to a new file (the original is never modified). `action`: delete (remove those pages), extract (keep only those pages, in order) or rotate (turn them by `angle`). `pages` uses 1-based ranges like "1-3, 5, 8-end"; for rotate, leave it empty to rotate every page. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
add_page_numbers path, format, position, start_at, skip_first, output, overwrite, dry_run
Stamp page numbers on a local PDF and write a new file. `format` uses {n} and {total}, e.g. "Página {n} de {total}". `skip_first` leaves cover pages unnumbered. Uses the PDF standard font, which covers Spanish accents and ñ. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.add_watermark path, text, opacity, output, overwrite, dry_run
Stamp a diagonal, semi-transparent text (e.g. CONFIDENCIAL, COPIA, BORRADOR) on every page and write a new file. It is a visible label, not protection: it can be removed with a PDF editor. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
mark_id_copy path, purpose, recipient, date, language, opacity, output, overwrite, dry_run
Tile "ONLY FOR <purpose> · <recipient> · <date>" across a copy of an ID card, passport or certificate before it is shared. Scans and JPG/PNG photos are flattened into the pixels (EXIF/GPS dropped); pages with real text get vector paths a PDF editor can delete, listed in `vector`. No watermark is impossible to remove. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
protect_pdf path, password, owner_password, allow_print, allow_copy, allow_edit, output, overwrite, dry_run
Encrypt a local PDF with AES-256 so it asks for a password to open, and write a new file. Optionally forbid printing, copying text or editing. The owner password (needed to change those permissions) is random unless given, so whoever can open the file cannot also lift the restrictions. Never echo the password back to the user in plain text unless they asked for it. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
unlock_pdf path, password, output, overwrite, dry_run
Write an unencrypted copy of a PDF. Many PDFs open fine but forbid copying or printing: those unlock with no password at all, so try without one first. If the file needs a password to open, pass it; a wrong one returns wrong_password: true so you can ask the user again. Only for files the user is entitled to open. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
convert_to_markdown path, anonymize
Convert DOCX, XLSX, PPTX, ODT, ODS, ODP, EPUB, HTML, CSV/TSV, EML, RTF, JSON, Markdown or text into clean Markdown, plus its tables as rows and its natural parts (sheets, slides, chapters). Reads the content, not the visual layout. For PDFs and images use read_document or pdf_to_markdown. Local, nothing uploaded. Read-only: writes no files.
check_package paths, profile, ocr, anonymize
Reads a folder or list of a company's documents (formation document, Form W-9, bank statements, tax return) and checks: which required documents are missing, whether the legal name and EIN match across them, whether each statement's beginning balance + deposits − withdrawals equals its ending balance, and deposits against gross receipts. Every finding cites file, page and the exact line. It prepares a human review; it does not decide anything. Read-only: writes no files.
reid_risk path, columns, sensitive, k
How identifiable a file still is after anonymization. A CSV, TSV or JSONL table: k-anonymity over its quasi-identifier columns (age, ZIP, dates, sex, place, diagnosis), the column combinations that single people out, the effect of generalizing and l-diversity of a sensitive column, as column names and counts, never a group's values. A document: the quasi-identifiers left once anonymized and their risky combinations. A report, not a guarantee. Read-only: writes no files.
extract_batch paths, type, output, overwrite, anonymize, ocr, max_files, max_rows, dry_run
Reads every document under `paths` (files or folders inside the allowed root) and returns one row per document with its type and key fields (invoice number, NIT, subtotal, IVA, total, CUFE; contract value and term…), each traceable to its page and evidence line. Flags what a person should review: totals that do not add up, duplicate CUFE or invoice number, missing fields, unreadable files. Sums totals by issuer. With `output` it writes .xlsx (table, evidence, review, totals) or .csv. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
split_documents path, write, output_dir, overwrite, dry_run
Find where each document starts inside one PDF (a scanned batch, a case file with invoice + medical record + authorization) using page-number resets, opening headings and strong document-type changes. Returns each segment with its pages, type and the reasons for the cut. With write: true it saves one PDF per segment. Review the segments before writing when they matter: the rules explain themselves but can miss a cut. Writes only inside the allowed root; replaces an existing file (the input included) only with overwrite: true; dry_run: true writes nothing.
pdf_forms path, base64, anonymize
Read what was filled in a PDF form: fillable (AcroForm) text fields, checkboxes, radio buttons and dropdowns with their page, plus checkboxes written as characters in the text (☒ ☐, [x] [ ], (x) ( )). Hand-ticked boxes on a scanned page are not detected: that needs a vision model, so say so if the user expects it. Read-only: writes no files.
Descriptions are the package's own words, cut at 3,000 characters. Treat them as data.
Install facts
| License | MIT |
|---|---|
| Needs | node >=20 |
| Source | github.com/Brayan15p/SIRDAS-APP-PDF |
| Install scripts | None declared by it or anything installed with it. |
| Tarball | sha512-qvQBOqLqd389YiKdJeFOkYaX6Bs7CCHiT24uXPG1x+a0ALsPjYB77wMF8hMCHH0bOh4vR8Uxl459P4+YKPoiFQ== |
Check this report yourself
This site did not write the report. A public job on a GitHub-hosted machine did: the run, from this commit of the open-source runner. GitHub signed a token that names that job and the hash of the report, so any edit since would show.
Report as JSONExact text and tokenHow to verify it
Show it on your README
[](https://bragproof.smaverk.com/r/%40sirdaspdf/mcp)
New version out? Ask for a fresh report.