How to fill PDF forms with a browser AI agent using SimplePDF

Benjamin André-Micolon

AI agents now live inside the browser: the built-in browser of the ChatGPT desktop app and Chrome can read the page you are on and act on it. Most sites leave those agents to guess their way through buttons and text boxes. SimplePDF does not. The editor tells the agent exactly what it can do with the open document through WebMCP, a web standard in the making (a W3C Community Group draft) for exposing a site's actions to AI agents, so the agent fills the form the way a script would, and you review the result before it goes anywhere.

Nothing to install or configure on the SimplePDF side. Open the editor in an agent-capable browser and ask.

#Who this helps

Letting an agent fill a form is not a party trick. For many people it removes a barrier that a mouse and a keyboard never will:

• People who cannot see the form. A blind or low-vision user drives the agent by voice: "read me this form", "put my name in the first field", "check the box for a single applicant". The agent reads the document through get_document_content, lists the fields, fills them, and reads back what it entered. The PDF becomes a conversation instead of a grid of unlabeled boxes a screen reader struggles with.
• People filling a form in a language they do not read. A rental contract in Dutch, a tax form in German, a school enrollment in Japanese: the agent reads the document in its language, explains each field in yours, and fills in your answers in the form's language. You still see every value on the page before anything is sent.
• Anyone facing a dense contract. Ask the agent to explain the document in plain words before you sign it: what you commit to, what the other side commits to, which clauses are unusual. The agent has the full text through get_document_content, so the explanation is about this document, not a generic template.
• People with motor impairments, or simply a 40-field form. Say the answers once. The agent places each one, and you fix the odd mistake instead of typing everything.

In every case the same rule holds: the agent proposes, you review. On the free editor the agent cannot submit for you: the submit tool answers with a sign-up error, so finishing a shared form stays with a signed-in person. Download is a tool too, so tell the agent to leave it to you (the ready-made instructions on the home page do exactly that), and check every value before the document goes anywhere.

#Which agents work today

WebMCP is being standardized at the W3C with Google and Microsoft behind it, and browser support is arriving in stages:

• ChatGPT desktop app, built-in browser: works today. ChatGPT discovers SimplePDF's tools as site tools (listed from the address bar) and uses them when you ask it to work on the document. Site tools need GPT-5.6 Sol or Terra and are not available in Enterprise or Edu workspaces (OpenAI's site tools documentation).
• Chrome 149 and later: available through the WebMCP origin trial. To try it yourself, enable chrome://flags/#enable-webmcp-testing, relaunch, and install the Model Context Tool Inspector extension to see and call the tools. Gemini in Chrome support is announced.
• Other browsers: nothing yet. The editor works exactly as before; the tools are simply not offered.

#Fill a form with the agent

1. Open the SimplePDF editor in an agent-capable browser and open your PDF from your device or from a link
2. Ask the agent what you need, for example: "Fill this form with my details: name Jane Doe, born 12 March 1990, email jane@example.com, and tick the newsletter box"
3. Watch the agent work on the page: it lists the fields, reads the document when it needs context, then sets each value
4. Review every value, correct what you want, then download or submit the document yourself

Give the agent the data in your prompt, or point it at a page it can read, such as a previous email or a profile page open in another tab. Site tools cannot read files on your computer.

#What the agent can do

Every tool maps to something you can do yourself in the editor. The agent sees the same document you see, and every change shows up on the page as it happens.

In the SimplePDF editor, the tool names carry a simplepdf_embed_ prefix, so they never clash with another site's tools.

#What stays private

While you edit, the PDF is processed in your browser, like every document you edit with SimplePDF: a file you open from your device never reaches a SimplePDF server, and a PDF opened from a link is downloaded straight from that link (SimplePDF relays it only when the link's host blocks browsers from fetching it directly). When you submit a form, the filled document is saved where the form's owner configured it, their own storage or SimplePDF's, exactly as it would be without an agent. What the agent reads through the tools, such as the list of fields, their values, and the extracted text, goes to the agent's model provider, exactly like text you would paste into a chat. Treat the agent the way you would treat anyone you show the filled document to.

Nothing is final until the document is downloaded or submitted. On the free editor the agent cannot submit at all (the tool asks for a sign-up), and you can ask it to leave download to you as well: review the fields, then finish. The agent's field edits are regular edits: undo them, change them, or close the tab.

#Turn it on for your own editor

WebMCP is on in the free SimplePDF editor. On your own branded editor it stays off until you enable it from your dashboard's editor configuration, available on the Pro plan.

Embedding the editor in your app? Turn on the Embed SDK's WebMCP option: it registers the editor's tools on your page and forwards each call into the editor's iframe. That is the only place agents find them: the embedded editor registers no tools of its own, and ChatGPT's browser never looks inside an iframe anyway:

import { EmbedPDF } from '@simplepdf/react-embed-pdf';

<EmbedPDF
  mode="inline"
  companyIdentifier="acme"
  document={{ url: 'https://example.com/form.pdf' }}
  // Keep the decision with the person: withhold submit
  webMCP={{ enabled: true, exclude: ['submit'] }}
/>

import { createEmbed } from '@simplepdf/embed';

createEmbed({
  target: '#editor',
  companyIdentifier: 'acme',
  document: { url: 'https://example.com/form.pdf' },
  // Keep the decision with the person: withhold submit
  webMCP: { enabled: true, exclude: ['submit'] },
});

webMCP: { enabled: true } registers every operation the SDK exposes as a tool, loadDocument included; exclude takes SDK method names ('submit', 'loadDocument', 'deletePages', …) and withholds the ones you want a person to keep. The full option reference is in the Embed SDK documentation.

For agents that do not run in a browser, such as Claude, Cursor, or a backend automation, use the remote MCP server or the prefill API instead: see how to pre-fill PDF forms with AI agents.

#No agent in your browser? Use SimplePDF Copilot

WebMCP needs an agent-capable browser. If you do not have one, or you would rather have the assistant built into the page than into the browser, SimplePDF Copilot is the same idea as a chat panel next to the editor, and it works in any browser:

• Bring your own AI, today. Open copilot.simplepdf.com, connect the AI you already use (Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek) or a model running on your own machine (Ollama, LM Studio), and chat with the form. The PDF stays in the browser; the AI traffic goes to the provider you chose.
• A fully integrated experience in your own app. Copilot is an MIT-licensed reference implementation: fork it, wire it to your AI provider, your tools and your storage, tweak the prompts and the UI, and ship it under your own domain (the embedded editor needs the Pro plan to run there). See how to self-host SimplePDF Copilot.

What SimplePDF Copilot is covers the review workflow and the data flow in detail.

That's it! The agent in your browser fills the PDF, and you review and submit it.

If you have any questions, feel free to reach out to support@simplepdf.com