Developer docs

Claude connector

Reference for the Redact PDF AI connector in Claude: add it, sign in, upload links, download links, restore_names, tools and limits of the remote MCP server.

Claude connector

Redact PDF AI is listed in Claude's connector directory. The connector is a remote MCP server: you sign in once with your redact-pdf.ai account, then ask Claude to redact or pseudonymize a document. No API key, nothing to install.

  • Server URL: https://mcp.redact-pdf.ai/mcp
  • Directory listing: claude.ai/directory/redact-pdf-ai
  • Sign-in: OAuth 2.1 with your redact-pdf.ai account. The connector acts only on that account and uses its page credits.

This page is the technical reference. For a plain-language, step-by-step walkthrough, see the guide for Claude users.

Redact a loan agreement in an AI chat: only names and organizations, always the date of birth and the reference number, never the lender.

Add it

claude.ai and the Claude apps: open the directory listing and connect, or add a custom connector in Settings → Connectors with the server URL above. Sign in when Claude asks.

Claude Code:

claude mcp add --transport http redact-pdf https://mcp.redact-pdf.ai/mcp

Then run /mcp in Claude Code and sign in.

Other MCP clients: any client that supports remote (streamable HTTP) servers with OAuth sign-in takes the same URL. The server publishes its OAuth metadata at /.well-known/oauth-protected-resource/mcp. An API key also works, sent as X-API-Key or as a bearer token starting with rdp_live_.

try_demo works before sign-in, so you can check the connection first.

How a file reaches the connector

A remote server cannot read your disk, and Claude does not pass chat attachments to remote tools. So the connector gets the file one of two ways:

  1. Upload link (the default). create_upload_link returns a private page on www.redact-pdf.ai, valid for 30 minutes, for one PDF up to 50 MB or one JPEG or PNG up to 10 MB. You choose the file there; it goes from your browser straight to our storage and never passes through the chat. Claude then starts the redaction with the upload_id. One link makes one job: a second file needs a new link.
  2. Small files inline. Files under 1 MB can be passed as file_base64.

PDF, JPEG and PNG are accepted, scans and phone photos included.

What comes back

  • Redacted PDF: a download_url to a page with a Download button, valid for 15 minutes. A plain request (a link preview, a fetch tool) never receives the file; only a click does.
  • Pseudonymize mode: the document as text with tags such as [Person 1] in place of names (document_text), and a correspondence_download_url for the tag → original table. Long texts come back as their start plus a signed document_text_url.
  • Counts per category. Results carry document ids, never your file names.

Links are bound to one document and its owner, and re-check ownership, retention and the account balance when they are opened.

Pseudonymize, work, put the names back

Pseudonymize a document, let the assistant work on the tagged text, then put the real names back on our side.
  1. Ask for a pseudonymized copy. Claude works on document_text, where names are tags.
  2. Claude writes its answer (a summary, a letter) and keeps the tags as written.
  3. restore_names takes the document_id, the text_version that came with the text, and Claude's text (up to 500,000 characters). We put the real names back on our side and return a 15-minute restored_download_url for you to open. Claude gets counts and any tags it could not restore, never a name.

If the document was edited or detected again in the Studio after the text was produced, the restore is refused (text_changed) instead of putting in names that a tag no longer stands for. The restored file is deleted after an hour, or at once with the document.

Tools

ToolWhat it does
create_upload_linkPrivate upload page for the file, valid 30 minutes.
redact_pdf_and_waitRedact or pseudonymize and wait for the result (up to 180 seconds, then it returns the job id to resume with).
redact_pdfStart a job and return immediately, for several documents in parallel.
get_job_statusCheck a job: uploaded → analyzing → redacting → redacted or error.
download_redactedA fresh download link for a finished document.
download_correspondenceA fresh link to the tag → original table of a pseudonymized document.
restore_namesPut the real names back into Claude's tagged text, on our side.
delete_jobDelete a document and everything kept from it now. Claude confirms with you first.
get_account_statusCheck which account is signed in.
try_demoRedact the first page of a file, or run a built-in sample, with no sign-in.

The redaction tools take the same options as the API: mode, pii_categories, pii_included_terms (always redact), pii_excluded_terms (never redact) and retention. See PII categories and Redaction rules.

Limits

LimitValue
Wait inside one tool call180 seconds, then resume with get_job_status
Calls per minute120 per signed-in user, 30 per address without sign-in
Redactions running at once2 per user
Inline file (file_base64)Under 1 MB; use an upload link for anything larger
Restored textUp to 500,000 characters in, 5 million out

A refused call comes back as a tool error with retry_after_seconds, so Claude can wait and retry.

Privacy

  • The file goes from your browser to our storage, not through the chat.
  • In redact mode Claude never sees the removed text.
  • In pseudonymize mode Claude sees the document with tags in place of names. Everything that is not a detected entity stays in that text.
  • Want nothing about the document to reach the AI? Use the local MCP server instead.

Details: Privacy policy, AI assistants section.

Related pages