Arama Yap Mesaj Submit
Request a Callback
+90
X
X

Select Your Currency

Turkish Lira $ US Dollar Euro
X
X

Select Your Currency

Turkish Lira $ US Dollar Euro

Contact Us

Location Halkali merkez neighborhood fatih st ozgur apt no 46 , Kucukcekmece , Istanbul , 34303 , TR
Last technical review · 17.08.2026 · Cloudflare WebMCP

Cloudflare WebMCP: Expose Secure Website Tools to AI Agents

WebMCP lets an agent discover explicit website tools instead of guessing UI interactions. This production guide covers discovery, invocation, permission boundaries, validation and rollback.

Production note

WebMCP is not a trusted hidden API layer. Treat browser-originated calls as untrusted, enforce authorization at the origin application and re-confirm sensitive actions.

webmcp guidecloudflare webmcpwebmcp setup
TECHNICAL IMPLEMENTATION PROFILE
EKA CORE
Cloudflare WebMCP

As of 6 August 2026 WebMCP is a developer preview. Start with a low-risk read-only tool; do not expose state-changing actions without user confirmation, strict parameter validation and server-side authorization.

06.08.2026Cloudflare launch
Checked
document.modelContextBrowser surface
Checked
Read → WriteRisk order
Checked
Origin AuthAuthorization boundary
Checked
Technical guide · production-focused · official sources
Quick answer

As of 6 August 2026 WebMCP is a developer preview. Start with a low-risk read-only tool; do not expose state-changing actions without user confirmation, strict parameter validation and server-side authorization.

01

Technical scope at a glance

Learn how WebMCP exposes structured website tools to browser AI agents, how to test discovery with Cloudflare Browser Run, and how to secure production usage.

06.08.2026Cloudflare launch

Cloudflare announced WebMCP support for Browser Run as a developer preview.

document.modelContextBrowser surface

Tools are declared through the page model-context surface and can be discovered by agents.

Read → WriteRisk order

A read-only pilot is safer; payment or account-changing tools should come later.

Origin AuthAuthorization boundary

Real authorization belongs at the origin application layer, not in the browser declaration.

On this page

  1. How a WebMCP call actually flows
  2. Design the first tool as a read-only order-status lookup
  3. Validate classic HTTP before involving an agent
  4. Threat model: separate agent mistakes from hostile input
  5. Release gate before production
  6. Frequently asked questions
02

How a WebMCP call actually flows

Tool discovery and secure execution are separate concerns. Treating the layers independently makes debugging and authorization clearer.

1) The page declares the tool name, description and parameter schema. The agent reads this contract instead of blindly clicking the DOM.
2) When the agent invokes a tool, the browser is only a bridge. Session, role, quota and business rules must be revalidated by the origin endpoint.
3) Return a concise structured result. Dumping HTML or full database rows increases token cost and data-leak risk.
03

Design the first tool as a read-only order-status lookup

For a pilot choose a narrow read-only action that returns the authenticated user’s own data.

Use a precise name such as `order_status`. Vague names like `do_action` increase tool-selection errors.
Constrain parameters with length, character-set and format validation for the order identifier.
The endpoint must verify that the order belongs to the authenticated session; the tool schema cannot guarantee ownership.
04

Validate classic HTTP before involving an agent

Many apparent WebMCP issues are actually origin, session or CORS failures. Isolate normal HTTP first.

Command
curl -I https://ornekalanadi.com
Command
curl -sS https://ornekalanadi.com/api/order-status -H "Accept: application/json"
Command
curl -sS -D- https://ornekalanadi.com/ | head -40
05

Threat model: separate agent mistakes from hostile input

Prompt injection, over-privilege and unvalidated parameters are different failure classes with different controls.

RiskSignalControl
Prompt injectionAgent selects an unexpected toolNarrow descriptions, allow-list, user confirmation
Over-privilegeWrite token used for a read-only taskSeparate service identity and minimum scope
Parameter abuseUnexpected ID, URL or file pathServer-side schema plus business-rule validation
06

Release gate before production

Version, log and roll back WebMCP tools like normal API changes.

Verify the tool output contains no unnecessary personal or operational fields.
Log successful and rejected calls separately; never log secrets or full authorization headers.
Keep a feature flag or configuration kill switch so the tool can be disabled instantly.
EKA SUNUCU · TECHNICAL

Build a secure server foundation for agent-ready web features

WebMCP depends on more than browser code: origin APIs, logging, TLS, rate limiting and access policy must work together. Eka Sunucu can size the infrastructure for this workload.

Production principleMeasure → Test → DeployNo fabricated benchmark data.
SRC

Official sources

Primary documentation and technical references used by this guide.

EKA

Related technical guides

Continue with related infrastructure and implementation guides.

FAQ

Frequently asked questions

Cloudflare WebMCP

Is WebMCP the same as classic MCP?

No. WebMCP focuses on exposing website tools inside a browser, while MCP is a broader client-server protocol for tools and context.

Can WebMCP be used in production?

A production pilot is possible, but preview status means higher change risk. Avoid making critical workflows depend solely on it.

Where should authorization happen?

At the origin application. The browser declaration is a contract, not an authorization decision.

What is a good first tool?

A low-risk read-only action such as status lookup, stock availability or documentation search.

Top