Give your AI assistant a real document pipeline
Tavnit runs an MCP server. Connect it to claude.ai or Cursor and your assistant can push documents through your own extraction flows and read the results back — returning the exact fields you defined, cleaned and typed, instead of an interpretation of a file it was handed.
Setup is a URL you paste into your assistant’s settings. No SDK, no server to run, no code.
Why assistants struggle with documents
AI assistants are good at reasoning and bad at being a data pipeline. Hand one a scanned invoice and it will read it — probably correctly, sometimes not, and differently the second time. There is no schema, so field names drift between runs. There is no validation, so a misread total looks exactly like a correct one. And there is no record of what happened, which matters as soon as the output touches accounting or compliance.
That is fine for a one-off question and unworkable as a process. What an assistant actually needs is a tool that already knows how to read your documents.
Pasting a PDF vs. calling a flow
| Capability | Pasting the file into chat | Through the MCP connector |
|---|---|---|
| Output shape | Whatever the model returns that time | The fields your flow defines, every time |
| Field names | Drift between runs | Fixed by your schema |
| Cleaning rules | None | Cleaners applied automatically |
| Human review | Not possible | Optional, with an audit trail |
| Where results land | In the chat | Buckets, webhook, email or API |
| Repeatable at volume | No | Yes — same flow, any number of documents |
What your assistant can do once connected
Run a document through a flow
Ask the assistant to process an invoice, contract or form through one of your extraction flows. It returns the fields you defined, typed and cleaned, not a paragraph describing them.
Query what you have already extracted
Ask questions against your Buckets — what you paid a vendor last quarter, which contracts renew next month — and the answer comes from your live data rather than the model's memory.
Things you can ask it:
- “Run this invoice through my Supplier Invoices flow and show me the line items.”
- “What did we pay Acme Corp last quarter, according to my Invoices bucket?”
- “Extract the renewal dates from these three contracts and compare them.”
- “Which runs are waiting on review right now?”
It uses the pipeline you already built
The connector is not a separate extraction engine. It calls the same flows as the rest of Tavnit, which means every rule you configured still applies when an assistant is the one making the request.
- Collections still classify an incoming document and route it to the right flow.
- Cleaners still standardise formats, convert currencies and apply your lookups.
- Human-in-the-Loop review still pauses a run when you have asked it to — an assistant cannot skip your approval step.
- Buckets still receive the results, so what the assistant extracts is queryable afterwards.
Scope and access
Your permissions, not more
The connector is issued from your own API key and reaches only the organisation you generated it in. The assistant inherits your role — it cannot do anything you could not do in the app.
Treat the URL as a credential
Anyone holding the connector URL can reach your flows and Buckets. URLs are time-limited, and refreshing one invalidates the previous value immediately.
Full detail in the connector documentation and user roles reference. The server endpoint is https://mcp.tavnit.io.
Connecting it
Generate a connector URL on the Integrations page in Tavnit, then paste it into claude.ai under Settings → Connectors, or add it to Cursor as a remote MCP server. It takes a couple of minutes and needs no code.
Read the step-by-step setup guideCommon questions
- How is this different from pasting a PDF into Claude?
- Pasting a PDF asks the assistant to read a document it has never seen, in a format it was not designed to parse, with no schema and no way to check the result. The MCP connector routes the document through an extraction flow you defined, so you get the same named fields every time, in the same types, with the same cleaning rules and the same review step applied. The assistant receives structured data rather than an interpretation.
- Which AI assistants work with it?
- claude.ai on a Pro plan or above, Cursor, and any client that accepts a remote MCP server URL. MCP is an open protocol, so support is not limited to a specific vendor.
- Do I need to write code?
- No. You generate a connector URL in Tavnit and paste it into your assistant's settings. Building the flow it calls is also no-code — you describe the fields you want and Tavnit extracts them.
- Can the assistant see everything in my company's account?
- No. The connector is issued from your own API key, so it reaches only the organisation you were signed in to, with your role's permissions. A Member cannot make an assistant do something a Member cannot do in the app, and private Buckets stay private unless you have been granted access.
- Does using the connector cost extra?
- No. Documents processed through the connector consume credits at the normal rate of one credit per page, the same as uploading through the app or calling the API. There is no separate charge for MCP access.
- What happens when the connector URL expires?
- Connector URLs are time-limited, and Tavnit shows when yours was created and when it expires. Refreshing issues a new URL and invalidates the previous one immediately, so any client still holding the old one stops working until you paste the new value.
Point your assistant at real documents
Build a flow, generate a connector URL, and start asking. Free credits to begin, and no card required.