Add Unstructured Transform MCP integration

[dizon-u10d]

Adds an integration catalog page for Unstructured Transform MCP, a hosted MCP server exposing Unstructured's document-processing pipeline (partition, enrich, chunk, embed) as MCP tools.

• New: integrations/transform-mcp.md (type: Tool Integration)
• Connects via the existing mcp-haystack MCPToolset over Streamable HTTP using StreamableHttpServerInfo's native token parameter; documents the mcp-remote OAuth fallback.
• Reuses the existing /logos/unstructured.svg asset.

Docs: https://docs.unstructured.io/transform/overview · Get started: https://transform.unstructured.io/get-started

Companion cookbook PR: deepset-ai/haystack-cookbook#300

🤖 Generated with Claude Code

[vercel]

@dizon-u10d is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[dizon-u10d]

Hi @kacperlukawski, following up on this listing for Unstructured Transform MCP (uses the existing mcp-haystack integration, same as integrations/mcp.md).

Verified against the live server: StreamableHttpServerInfo/MCPToolset correctly reaches https://mcp.transform.unstructured.io, and the server returns a clean 401 invalid_token with a spec-compliant OAuth-discovery challenge. One thing worth flagging: with eager_connect=True, an invalid/expired key doesn't surface that 401 right away, it hangs for mcp-haystack's ~30s connection timeout and then raises a generic MCPConnectionError, since the underlying mcp 1.28.1 SDK's streamable-http transport doesn't propagate the HTTP error out of its TaskGroup until the timeout fires. Not a defect in this doc, just something a user hitting a bad key would notice.

The only failing check here is the Vercel preview needing fork authorization. Happy to make any changes.

[Copilot]

Pull request overview

Adds a new integration catalog page documenting how to use mcp-haystack to connect Haystack agents to Unstructured’s hosted Transform MCP server for document processing.

Changes:

• Added a new integration page for Unstructured Transform MCP (overview, install, auth, usage, and agent example).
• Documented Streamable HTTP connection via StreamableHttpServerInfo(token=...) and an mcp-remote bridging fallback.
• Included licensing/terms notes for both the mcp-haystack client and the hosted Unstructured service.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[dizon-u10d]

Updated for the recent Transform MCP tool rename: transform_files→start_transform_job, check_transform_status→check_job_status, get_transform_results→get_job_results. Ready for another look whenever you get a chance.

[kacperlukawski]

Hey, @dizon-u10d I have some questions about that integration. I like the idea, but want to make sure the example is easy to follow. Would love to hear your thoughts about my concerns.

[dizon-u10d]

Hi @kacperlukawski, thanks for the review, both good catches. Pushed a fix for each:

• Dropped the mcp-remote fallback paragraph/snippet since it's not needed for this integration, the token-auth path already documented above is what actually works here.
• Added an explicit "waiting a few seconds between calls" to the polling step in the agent's system prompt, so it doesn't hammer check_job_status in a tight loop.

Let me know if anything else stands out.

[dizon-u10d]

Another round of tool-surface changes shipped upstream this week (file manager tool added), so I went further than a rename this time: rewrote the docs, notebook, and system/agent prompts to describe the pipeline by behavior (submit → poll → fetch) rather than hardcoding exact tool names, and lean on the toolset-discovery pattern already shown here to find the real names at connect time. Should hold up through future renames/additions without needing another update like this.

[kacperlukawski]

Thank you @dizon-u10d I still have some concerns about too many poll operations. Did you try running the code samples? I don't think we can just ask model to sleep, but a simple tool should work.

[kacperlukawski]

Suggested change

	toc: true
	---
	toc: true
	mcp: true
	---

We have an mcp property to distinguish MCP-based integrations from native ones.

[kacperlukawski]

Unfortunately, asking the model to wait a few seconds is not going to work. Chat generators are not able to control their execution, but it should be solvable with a simple tool.