Unstructured Transform

Parse PDFs, scans, spreadsheets and email into structured AI-ready data.

Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.

Paste it into any MCP client. Setup docs

What the Unstructured Transform connector does

The Unstructured Transform MCP server turns documents into structured, AI-ready data from inside an agent session. Point Claude, ChatGPT or any MCP client at a PDF, scan, spreadsheet or email and Transform partitions, enriches, chunks and embeds it, returning Markdown, JSON, HTML or plain text. Six tools, asynchronous jobs, OAuth sign-in, 15,000 free pages a month.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Unstructured Transform tools (6)

  • check_job_status
  • get_job_results
  • request_file_upload_url
  • start_extraction_job
  • start_transform_job
  • suggest_extraction_schema_for_file

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • The OAuth scopes are identity claims, not permissions. The server's RFC 9728 descriptor and its authorization server metadata both advertise exactly email, profile and offline_access. Nothing at the grant separates polling a job from starting one that spends your page allowance. There is no read-only mode.
  • We could not read tool schemas or annotations first-hand. The endpoint returned 401 to an anonymous request. Parameter detail on this page comes from the Smithery registry listing and Unstructured's documentation, not from a first-party handshake, and no readOnlyHint or destructiveHint value is reported anywhere.
  • A registry lists a seventh tool the vendor does not. get_instructions appears in the Smithery listing and in neither Anthropic's directory nor Unstructured's own repository. Six is what both first-party sources say; seven is what one third-party observation shows.
  • Free-tier work stops rather than degrades. Unstructured states that after 15,000 pages in a calendar month, processing requests begin failing unless you have upgraded, and that you must wait for the first of the next month otherwise. You are warned at 13,500 pages.
  • Wrong partitioning degrades silently. Unstructured's own tool description warns that vlm and hi_res do not apply to formats such as DOCX, XLSX, HTML and CSV, "silently fall back to fast", and that the wrong choice "degrades silently rather than erroring". A poor parse produces a sparse extraction with no error to catch.
  • Enrichments require hi_res. Unstructured documents that enrichment types need the hi_res partitioning strategy to have elements to act on; the fast strategy does not support them.
  • Embedding locks you to a model. The tool description states that the provider and model in an embed stage define the vector space, and that queries must be embedded with the same provider and model to be comparable. Record what you used.
  • Extraction never reads a raw file. Both extraction tools require Unstructured Element JSON, so a parse job is a mandatory first step, and embedded output is rejected as extraction input.
  • The connector holds no memory of your files. There is no search or listing tool. Persistent, indexed access to an organisation's document sources is a different Unstructured product — Foundation — which Unstructured documents as a separate, Early Access MCP server, not this one.
  • Retention is documented for dedicated instances, not for this connector. The ephemeral-use statement we quote sits in Unstructured's dedicated-instances security documentation. We found no Transform-specific retention page and did not test it.

Frequently asked questions

How many tools does the Unstructured Transform MCP server have?

Six by Anthropic's count and Unstructured's own, and a registry listing shows seven. Anthropic's directory and Unstructured's public integrations repository name the same six job tools. The Smithery registry lists those six plus get_instructions, a self-describing tool that returns the server's usage guidance. Treat six as the documented floor.

Can the Unstructured Transform connector send my documents to another system?

No tool in the listing writes to an external destination. The six tools upload a file, run a parsing job, poll it and fetch results back to you. Unstructured's wider Platform API does have destination connectors and a run-workflow endpoint, but none of those appear as MCP tools, so an agent cannot configure or trigger a delivery pipeline through this connector.

What OAuth permissions does the Unstructured Transform MCP server request?

Only identity claims. The server's RFC 9728 descriptor advertised email, profile and offline_access on 2026-08-22, and its authorization server metadata advertises the same three. None of them separates parsing a document from spending your page allowance, so consent is all-or-nothing across the six tools.

Do the results come back inside the conversation?

No, they arrive out of band. The get_job_results tool returns a short-lived signed download URL and element counts rather than the document text, so a large parse never floods the agent's context. Unstructured's tool description tells the agent to hand you a curl command instead of reading the output back into context.

Does Unstructured Transform cost money to use?

The first 15,000 pages each month are free. Unstructured's billing page states pages 15,001 to 115,000 are billed at $0.03 per page capped at $3,000, pages up to one million are not billed, and billing resumes past a million. Without a Pay-As-You-Go upgrade, requests start failing once the free 15,000 pages are used.

What are the file limits on the Transform MCP server?

Fifty megabytes per file, ten files per request and five jobs running at once. Unstructured's Claude Code setup guide lists all four limits, adding that each file must be a supported type. The server is designed to tell your agent about these caps, so Claude should batch requests rather than fail on an oversized file.

Does embedding a document send it to OpenAI or another AI provider?

Yes, when you ask for enrichment or embedding. The embed stage defaults to OpenAI's text-embedding-3-small, and enrichments such as image descriptions and generative OCR are vision-model passes. Unstructured's tool description names Anthropic, OpenAI, Bedrock and Vertex AI as selectable providers, so document content reaches whichever provider you pick.

Does Unstructured keep my documents after parsing them?

Not long-term, by Unstructured's own statement. Its security documentation says Unstructured does not store customer data long-term, that customer artifacts are used ephemerally during workflow execution and are removed on completion or failure, and that retained telemetry does not contain file data. That page describes dedicated instances, so confirm it covers your plan.

Why does Claude Desktop fail to reach the Transform server?

Because the domain is not on your network allowlist. Unstructured's setup guide instructs Pro and Max users to switch on code execution and network egress, then add mcp.transform.unstructured.io to the allowed domains list. Free plan users cannot allowlist the domain at all and must follow the separate no-domain-access path in the same guide.

Sources

Use it in an agent

Put Unstructured Transform to work.

Connect Unstructured Transform once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.