Use Novus PDF Studio From an AI Client

If your assistant supports the Model Context Protocol, you can add this site to it as a server. What that gets you is a straight answer to which tool does a job, what it accepts, what it will not do, and the exact address that opens it. What it does not get you, and cannot, is the assistant doing the work on your document.

The second half of that is the more useful thing to understand, so it goes first.

Every tool on this site runs inside the visitor's browser tab. There is no upload endpoint and no server-side document store, which is the claim the rest of the site rests on. The consequence for an AI client is direct: the server at the other end of that connection holds no PDF of yours, cannot be handed one, and therefore cannot merge, split, sign, protect or export anything at all.

A tool named merge_pdf would have to do one of two things: ask you to upload a document to a place with nowhere to put it, or report a result it never produced. Neither is acceptable, so neither exists. What the server does expose is what it genuinely holds: the tool registry, the written steps for operating each surface, the limitations each one states, and the URL that opens it.

The endpoint is /mcp. One stateless POST, JSON-RPC over HTTP, protocol revision 2026-07-28. A GET answers 405 on purpose, because that revision removed the server-sent stream and a client opening one deserves a definitive answer rather than a request that hangs. The MCP server page carries the address, the configuration block most clients take, and the current tool list rendered from the same definition the endpoint dispatches on.

There is no account, no key and no sign-up. Everything the server returns is already public on the site's own pages, so an authentication step would be theatre: a caller reaches exactly what an anonymous visitor reaches, and nothing beyond it.

• `list_capabilities` returns every single-purpose tool plus the three workspaces, with what each accepts, what it produces, and the URL that opens it. It can be narrowed to one job group, or to the tools that take a particular kind of input.
• `find_tool` takes the job in ordinary words, such as putting two contracts into one file or fixing a scan that came out sideways, and returns ranked matches with their addresses.
• `open_tool` takes one identifier and returns the address plus everything written about that surface: what it accepts and produces, its privacy note, the authored steps for operating it, the tips, the limitations it states, its answers to common questions and its related tools.
• `search_docs` searches the written material rather than the tools: guides, tutorials, tips, direct answers, how-it-works steps, articles, and the product and policy pages, with deep links to the section named.
• `list_novus_sites` returns the rest of the network for a request this site does not cover, each entry with its own endpoint.

There is also one prompt, pick_a_tool, which walks a client through choosing a surface for a described task and reporting back what it will not do for that task as well as what it will.

1. 1

   Describe the job

   Tell the assistant what you are trying to achieve rather than which tool you think you want. The search is built for plain description and returns several candidates instead of one guess.

2. 2

   Get the operating detail

   Opening the best match returns its steps and its limitations. This is where you find out that a tool will not do the thing you assumed, while changing plan is still cheap.

3. 3

   Open the address

   The URL is the handoff. A browser does the work from that point on: an AI browser driving the real page, or you driving it yourself.

4. 4

   Supply the file yourself

   The document is chosen in the browser, from your own disk, and stays there. Nothing about the conversation that led you to the page changes that.

The server's own instructions single out two claims, because a summary is exactly where an overstatement gets made. The erase tool's default mode covers content rather than deleting it, so a covered paragraph is still inside the file. And a signature drawn here is an electronic signature, not a certificate-backed digital one. An assistant condensing a tool list has every incentive to round both of those up, and the instructions exist to stop it.

If either distinction matters for what you are doing, the long versions are worth reading before you act on a summary: electronic against digital signatures covers the first, and redacting sensitive information covers the second.

• Not a processing interface. No call accepts a document. That is a design decision rather than an omission, and nothing is planned that would change it.
• Not authenticated, because there is nothing private behind it to protect. The catalogue it serves is the catalogue on the pages.
• Not a write surface. Nothing a caller does through it changes anything on this site.
• Not unlimited. Requests are rate limited per client, at a level one conversation's worth of calls will not come near.

If none of that suits what you are trying to do, the ordinary route is still open. The tools hub is the same catalogue in a form a person reads, and the file never leaves your machine either way.