Page tools for AI agents (WebMCP)
A page can list actions an AI agent in the browser may call, instead of the agent guessing at buttons.
WebMCP is a proposed web standard. A page registers named tools with a description and input schema through navigator.modelContext, or marks up forms. A browser agent can then call a tool such as search flights or add to cart directly. It is an early preview in Chrome behind a flag and may change.
What it can and can't do
What the user does: None for the page owner. Agents or browsers may ask you before they call a tool
It can
- Expose named actions with typed inputs
- Let an agent skip fragile clicking
- Keep your own checks and limits in your tool code
- Work next to the normal page for people
It cannot
- Run without a browser agent present
- Stop an agent from using the visible page
- Prove who the agent is
- Ship to all browsers today
How it differs between platforms: Only Chrome has an early preview, behind a flag. Names, shapes and permission rules can still change.
Browser support
Version notes
- Chrome: Early preview behind a flag; may change
- Edge: No public status found
- Firefox: No public status found
- Safari: No public status found
How a site asks for it
| API | navigator.modelContext.registerTool({name, description, inputSchema, execute}); declarative form attributes |
| How access is granted | none |
| What the user sees | None defined yet in the draft |
| Needs a secure context (HTTPS) | Yes |
| Needs a user gesture | No |
| How long it lasts | Tools live while the page is open |
| How to take it back | Close the page or unregister the tool |
| Inside an iframe | Draft rules for frames are not settled |
Typical flow
- Page registers tools with names and input schemas
- An agent in the browser reads the tool list
- The agent calls a tool; the page code runs it
- Browser rules for who may call tools are still being drafted
Data you get
Shape: object
How often: set by the page
| Field | Type | Unit | Range | Precision |
|---|---|---|---|---|
name | string | None | tool name | Not listed |
description | string | None | what the tool does, read by the agent | Not listed |
inputSchema | object | None | JSON Schema for the arguments | Not listed |
execute | function | None | page code that runs when the agent calls the tool | Not listed |
Technical limits
- Early preview only
- Permission model not final
- Not identity; says nothing about who the agent is
The story
What it is
A site tells agents what they can do, in a menu, rather than letting them poke at the screen.
For you
You will not see it. If you use a browser agent, the sites it visits may offer tools so the agent works faster and makes fewer mistakes.
For makers
Validate every argument in your tool code and apply the same rules as for a user. Do not treat an agent call as safe. Plan for the spec to change.
Where you see it
| Type | Example | Apps |
|---|---|---|
| shopping | A store exposes add to cart and check stock as tools. | None named |
| travel | A travel site exposes a search flights tool that returns structured results. | None named |
| productivity | A task app exposes create task and list tasks. | None named |
| developer | A team tests how agents use their tools in Chrome with the flag on. | Chrome |
| enterprise | An internal portal publishes approved actions for staff agents. | None named |
| accessibility | Clear tool names help agents act for people who cannot use a mouse. | None named |
Risks
- Prompt injection through tool descriptions
- Agents acting without a clear person's yes
- Preview may change
What the platform says
- W3C Web Machine Learning Community Group: WebMCP. Draft spec for exposing page tools to agents through the browser.
Good uses and abuse
Legitimate uses
- Agent friendly sites
- Faster, safer agent actions than clicking
Known abuse
- Prompt injection
- Agents doing more than the person meant
Scores
| Score | Value | Meaning |
|---|---|---|
| Sensitivity | 3 of 5 | How revealing the data is |
| Utility | 4 of 5 | How useful it is to honest apps |
| Friction | 1 of 5 | How much the user has to do |
| Fingerprint bits | 0.1 | Rough identifying entropy, when known |
History
- 2026-02, Chrome 146: WebMCP early preview in Chrome behind a flag
Related capabilities
Related
- Automation (webdriver) flag. Can tell if Selenium, Playwright, or Puppeteer is controlling the browser
- Bot detection and CAPTCHA signals. Can combine many browser and network clues to guess if you are a bot
- Passkeys and WebAuthn sign in. Can sign you in with a passkey after your fingerprint, face or PIN
- Signed agent requests (Web Bot Auth). Can prove which agent operator made a request
Sources
How well checked: Inferred from related behavior.