Unstructured Transform — ClawHub bundle plugin

Production-grade document processing your AI agents can call as MCP tools. This bundle registers Unstructured's publicly-hosted Transform MCP server with an OpenClaw (or Claude Code) agent. Parse PDFs, invoices, spreadsheets, presentations, scanned docs, and 60+ file types into reliable structured output — markdown, Element JSON, HTML, or plain text — with tables and layout intact.

What it wires up

Tools exposed (7): request_file_upload_url, start_transform_job, check_job_status, get_job_results, suggest_extraction_schema_for_file, start_extraction_job, get_instructions.

What you can do

• Extract clean, structured text from PDFs, invoices, spreadsheets, presentations, scanned docs, emails, HTML, and images — 60+ file types.
• Get output as markdown, Element JSON, HTML, or plain text, ready for RAG, search indexing, or downstream agents.
• Configure a partition → enrich → chunk → embed pipeline: pick a partition strategy (fast / hi_res / vlm / auto) and enrichments (table_to_html, image_description, generative_ocr, ner), and optionally chunk + embed for RAG ingestion.
• Preserve tables (returned as embedded HTML) and OCR scanned/image PDFs.
• Extract structured JSON against a schema — draft one with suggest_extraction_schema_for_file, then run start_extraction_job.
• Handle large, multi-page documents via an async submit-and-poll job flow.
• Parse from a public URL directly, or upload a local file first.

Package contents

.mcp.json                                # MCP server registration (Claude Code format)
.claude-plugin/plugin.json               # Claude Code plugin manifest
openclaw.plugin.json                     # OpenClaw bundle manifest (schemaVersion 2026.1)
package.json                             # openclaw compat metadata
skills/transform-unstructured/SKILL.md   # bundled skill: the upload→transform→poll→retrieve workflow

Auth

The server advertises OAuth 2.0 Protected Resource Metadata at /.well-known/oauth-protected-resource, so the client completes OAuth automatically on first connect. No secrets are stored in this bundle.

For headless use, present an Unstructured API key as a Bearer token instead (get one at https://unstructured.io/getapikey) by overriding the server entry in .mcp.json:

"headers": { "Authorization": "Bearer ${UNSTRUCTURED_API_KEY}" }

Install & test in OpenClaw

1. Install the plugin

openclaw plugins install clawhub:@unstructured/transform-mcp

This registers the transform-unstructured MCP server and the bundled skill. Restart / reload OpenClaw afterward so it picks up the new server.

2. Authenticate (browser OAuth)

openclaw mcp login transform-unstructured

A browser window opens — sign in and approve; tokens are stored automatically. Nothing to paste. (On a headless host, use an Unstructured API key as a Bearer token instead — see Auth above.)

3. Verify the connection

openclaw mcp doctor transform-unstructured --probe

A healthy probe lists the seven tools: request_file_upload_url, start_transform_job, check_job_status, get_job_results, suggest_extraction_schema_for_file, start_extraction_job, get_instructions. openclaw plugins list confirms the plugin and skill loaded.

4. Smoke test

Ask the agent to parse a public URL (no upload needed):

"Use the transform tools to parse https://<some-public-pdf-url> into markdown."

Expected chain: start_transform_job → job_id → poll check_job_status → get_job_results returns a download URL. Point it at a local file to also exercise the upload path (request_file_upload_url + PUT).

Exact openclaw subcommand names can vary by version; openclaw plugins --help and openclaw mcp --help confirm the current forms.

Publish (maintainers)

clawhub login
clawhub package publish . --family bundle-plugin --owner unstructured --dry-run
# then re-run without --dry-run