AI agents in your browser (WebMCP)
WebMCP is a new web standard that lets an AI agent built into your browser call tools a website publishes, instead of reading the screen and clicking around. OneBooks publishes such tools in the app, where an agent can look things up in your books and prepare documents for you to confirm, and on this documentation, the website and the developer portal, where it can read published information.
How it differs from connecting over MCP
Section titled “How it differs from connecting over MCP”| Connecting over MCP | WebMCP | |
|---|---|---|
| Where the agent runs | On the AI provider’s servers or on your computer — Claude, ChatGPT, Claude Code | In your browser, on the OneBooks page you have open |
| How it gets access | You approve access once on a consent screen | Your signed-in browser session; there is nothing to connect |
| Who confirms a change | You approve the preview in the chat, then the agent confirms it | You press Confirm in OneBooks itself |
Both use the same OneBooks tools under the same rule: nothing is written until a person approves it. To connect an agent from outside the browser, see Connect Claude, ChatGPT and other AI agents.
Browsers and agents that support it
Section titled “Browsers and agents that support it”As of October 2026:
- ChatGPT desktop app — its built-in browser uses WebMCP tools.
- Google Chrome and Microsoft Edge — through an origin trial, an early-access programme for new browser features.
- Brave — behind a flag, an experimental setting you turn on yourself.
- Gemini in Chrome — announced by Google.
- Claude in Chrome does not read WebMCP tools yet. If you use Claude, connect it over MCP instead: Connect Claude, ChatGPT and other AI agents.
In a browser without WebMCP nothing changes: OneBooks registers no tools and loads nothing extra.
In the OneBooks app
Section titled “In the OneBooks app”While you are signed in, the agent works as you: it has your role and permissions in your active business, exactly as if you were using the app yourself. A viewer gets read-only tools; every other role gets the tools the AI assistant offers that role.
What the agent can do
Section titled “What the agent can do”The agent always has a few general tools — your business details, a search of this documentation, the dashboard figures and the status of a proposal — plus the tools for the area on screen: sales, purchases, banking, accounting, reports, items or budgets. It can also:
- tell which page you are on (
get_current_page); - open another page (
open_page), which brings that area’s tools with it; - list what is waiting for your approval (
list_pending_approvals).
It cannot open the sign-in and consent pages, onboarding, the Apps section, or the settings that manage users, roles, API access and billing. As everywhere in OneBooks AI, there are no delete, void or refund tools.
Agent approvals
Section titled “Agent approvals”When the agent prepares something — an invoice, a payment, a journal entry, an edit — OneBooks stores it as a proposal and saves nothing yet. The Agent approvals panel opens with the proposal in it, a notification with a Review button announces it, and while anything is waiting, a button with the count stays in the header.
Each card shows the whole proposal, the same card the assistant uses, labelled Prepared by an AI agent in this browser. Press Confirm to save it or Cancel to drop it. A proposal nobody confirms expires after 24 hours, and at most 20 can wait for you at once.
Turn it on or off
Section titled “Turn it on or off”Open Settings → User Preferences. The AI agents in this browser card has the switch Let AI agents in this browser use OneBooks tools. The choice is stored in this browser only and is on by default wherever the browser supports WebMCP; switching it off removes the tools from the page at once and leaves your other browsers as they are. In a browser without WebMCP the card says so and the switch is unavailable.
On the documentation, website and developer portal
Section titled “On the documentation, website and developer portal”These sites publish read-only tools over the information they already show. They need no sign-in and see nothing of your books.
- This documentation (docs.getonebooks.com) — search the docs in the page’s language, read a page as Markdown, list the sections, describe the open page and open another one (
search_docs,get_doc_page,list_doc_sections,get_current_page,open_doc_page). - The website (getonebooks.com) — prices for a country in its billing currency, a plan comparison, features and integrations, a search of the Marketplace, the FAQ and the ways to contact OneBooks.
- The developer portal (developer.getonebooks.com) — search the API reference and read an operation or a schema, list the OAuth scopes and the webhook events, and search and read the developer docs.
Credits and limits
Section titled “Credits and limits”In the app, each tool call counts against your business’s AI credits like a call over MCP — 0.1 credit per call — and a weekly call cap per business applies, counted separately from MCP. The tools on the documentation, website and developer portal cost nothing.
Privacy and safety
Section titled “Privacy and safety”- The tools run in your browser with your session. The agent receives tool results — never your password or your session — and sees only what your role can see.
- Text stored in your records, such as a note or a name, reaches the agent marked as data, not as instructions.
- Tools on the documentation, website and developer portal read only published content.
- Nothing in your books is saved until you press Confirm, and you can turn the tools off at any time in Settings → User Preferences.