Dogfooding · WebMCP

Agent-first is the new
mobile-first.

Fifteen years ago, "does it work on a phone" became a basic question for every website. A new question is emerging: can an agent understand your product and use it on someone's behalf? We started with our own contact form, then turned what we learned into knowledge other teams can use.

The Shift

A new way to use your website

Mobile-first asked teams to design around how people actually used their phones. Agent-first asks a similar question: what does a customer need an agent to understand and do on your site?

A customer might ask an agent to request a quote, find a booking, or update an account. The agent needs context about your product, a clear way to perform the task, and feedback that tells it whether the action succeeded.

WebMCP is a proposed web standard that lets pages expose named tools to agents. Each tool describes its purpose and accepted inputs. It runs inside the page, where it can use the user's existing session and update the visible interface.

That gives teams a way to define actions directly. An agent can call a tool to fill a form rather than infer the task from screenshots, page structure, and button labels. A well-designed tool can reduce the context and interaction steps needed to finish a task; the benefit depends on the workflow and should be measured.

What We Shipped

We put it on our own contact form

Our contact form below exposes three WebMCP tools alongside the regular Send button. Filling the form and sending it are separate actions, and submission still requires Cloudflare Turnstile verification.

// fill_contact_form: populates the form, does not send // wait_for_contact_verification: waits for the Turnstile token // submit_contact_form: sends, same endpoint the button uses await document.modelContext.registerTool({ name: "fill_contact_form", description: "Fill the Paniolo contact form fields on this page...", inputSchema: { /* name, email, org, type, message */ }, execute: async (input) => fillContactForm(input), });

In our end-to-end test with a WebMCP-capable browser, the fields filled visibly, the Turnstile widget showed "Success," and the page confirmed "Message sent." The tools use the same submission endpoint and server-side verification as the human interface.

The important work was deciding what each tool should do, how to describe it, and what happens when validation or verification fails. Those decisions connect the knowledge an agent needs with the harness it acts through.

In a supported browser with site tools enabled, you can ask your agent to fill this form with details you provide, then review the fields before asking it to send. The badge indicates that the page has registered tools; completing the workflow also depends on your agent's support and verification succeeding.

The Paniolo Intelligence Layer

From knowledge to reliable action

An agent needs to understand your product and know how to act within it. Documentation supplies context; tools expose actions; permissions and feedback keep those actions grounded.

Paniolo builds the intelligence layer that connects these pieces. Our work spans agentic knowledge and harness engineering: organizing the knowledge agents draw on and building the environments they operate through.

The knowledge side turns your sources, domain expertise, and operating procedures into a knowledge base agents can retrieve and apply. The harness side supplies instructions, tools, constraints, and checks that help them carry out the work. Together, they give agents relevant context and a reliable way to act.

A website is one place these pieces meet. So is a coding environment. The customer agent using your product and the coding agent building it both need to understand the task, take an allowed action, and recognize the result.

Our contact form follows that same cycle. We implemented the tools, tested the workflow, and captured the lessons in a reusable skill. The implementation gives an agent a way to act; the skill gives other teams and their coding agents the knowledge to build one.

Where This Goes

Start with a workflow worth improving

WebMCP is still early. Chrome and Edge offer origin trials, which let participating sites test experimental browser features. ChatGPT Desktop supports site tools in its built-in browser for eligible accounts and models. Availability varies, and the API is still evolving.

The opportunity is to learn from a real task now. Choose a workflow customers care about, identify the knowledge it requires, expose the actions, and test completion with a supported agent. Measure success, failure, and the points where a person needs to review or intervene.

There is potential here for people who use agents as assistive intermediaries, too. Declared actions give those agents another way to interact with a site. They complement accessible interfaces; teams still need to test the experience with the people who will use it.

Get Started

Put knowledge to work

DIY

Build with our skill

Give your coding agent the knowledge we captured: tool design, when to use WebMCP or a backend MCP server, security considerations, and implementation failure modes. Install the free skill and use it to plan and test your first workflow.

npx skills add paniolo-ai/skills --skill paniolo-webmcp-dev
WITH US

Build your intelligence layer

Start with a customer workflow or a task your coding agents struggle with. We help organize the knowledge it requires, design the tools and permissions, and implement and verify the workflow. The scope can include your agentic knowledge base, your harness, or both.

Tell us about your workflow

Work With Us

Help agents understand
and act.

Tell us what your agents need to know, what they need to do, and where they get stuck. We help teams connect agentic knowledge bases, coding harnesses, and customer-facing tools into an intelligence layer built around their work.

Connect what your agents know with what they can do.

Agentic knowledge and harness engineering, built around your workflow.

Discuss your workflow