WebMCP area · diagnosis, training and projects
The agents in the area AGI‑ready need somewhere to act. Today they arrive at your site, try clicking, typing and guessing. With WebMCP, the page itself describes what it knows how to do, and the agent talks with capabilities instead of operating pixels. This area combines free diagnosis, mental model, architecture, and WebMCP Training in four phases.
Passive scanner: collects evidence without running any tool on your site.
01 · What’s changing
Three changes on the visitor side and three on the publisher side. If you own the site, the last three are your list.
Models like GPT‑6 Astra open pages, fill forms and complete tasks on behalf of someone. The Computer Use course shows five flows of this. Without site help, the agent works blind. Computer Use shows five flows of this. Without site help, the agent works blind.
More and more people ask an AI assistant instead of typing into a search engine. The responder is the model, with what it managed to extract from the pages. Appearing there is a new problem, called GEO and AEO, and classic SEO does not solve it alone.
WebMCP is an experimental technology, currently behind an Origin Trial in browsers. Those who understand the limitations early learn at a low cost. Those who wait for the standard to close arrive when competitors already have exposed tools.
In addition to the interface for people, the site now has a catalog of capabilities for agents: search, verify, initiate, confirm, consult. The human journey continues. The catalog sits on top.
Public files such as robots.txt, sitemap.xml and llms.txt, structured data in JSON‑LD, Open Graph and a clear identity of the entity. It is what an agent reads before deciding if your site is worth visiting.
Human confirmation before mutations, cancellation, fallback, permissions and real backend authorization. The rule of formation: no security decision is handed to the LLM.
On the agent‑command side, what changes in the work is the end of micro‑prompting and the architecture prompt + skills + context + tools + memory + evals, see the AGI‑ready area. This page is about the other side: the site that agents visit.
Open the AGI‑ready area →02 · What is WebMCP
Without WebMCP, an agent needs to locate fields, click, type and interpret the result on the screen. With WebMCP, the page itself describes capabilities, arguments and responses, and the agent calls the right tool by name.
Without WebMCP · the agent guesses
"I found a field that looks like a search. I typed the course name. I clicked the blue button. The page changed. I think it worked, but I don’t know if the spot was reserved.",
The agent operates the interface made for people. It works sometimes, breaks when the layout changes and no one knows what actually happened.
With WebMCP · the site describes itself
"The site exposesverificar_vagasandiniciar_inscricao, with the arguments and responses defined. I called the first, there is a spot. The second requires human confirmation before concluding.",
Tools with name, description, schema and understandable response. The site decides what is allowed, what needs confirmation and what can be cancelled.
The always‑present form gains a toolname and a description. Those who already have the human journey ready start here: little change, observable evidence in the scan.
The page registers tools with name, description, JSON Schema of input and output, handling state, errors, cancellation and fallback. It is the path to capabilities that don’t fit in a form.
Without WebMCP, the agent operates its screen.
With WebMCP, it converses with its site.
03 · Free diagnosis
The WebMCP Readiness opens the URL in a disposable browser, gathers observable signals and turns the scan into a correction plan. Consult robots.txt, sitemap.xml and llms.txt. It does not execute any site tool.
You receive overall score and four independent scores, approved evidence, alerts and blockers, up to 12 prioritized fixes and the courses recommended for each gap. The report comes out in JSON, with the evidence.
What the scanner can prove
toolname and tools returned by getTools()What requires human review
Honesty of the diagnosis. The scan only deepens the provided URL; the sitemap is inventoried, but its pages are not crawled in this mode. Advanced phase scanners (Builder, Integrator, Agent Developer and Expert) appear in the report as planned, not as ready.
04 · The architecture of a site ready for agents
Eight layers, in the order the training builds. The first four are the site describing itself. The last four are the site protecting itself.
Does the agent find the site?
robots.txt, sitemap.xml, llms.txt, JSON-LD, Open Graph, canonical and an entity with name, authorship and update.
What can the site do?
Declarative, from forms, and imperative, registered in JavaScript. Few, with names that say what they do.
What is it called and what does it return?
Name, description, JSON Schema of input and comprehensible response. The description must match the real effect.
And when it goes wrong?
Modeled state, explained errors, possible recovery. The agent needs to know where it stopped.
Who confirms?
Human confirmation before mutations, cancellation and fallback to the human journey when the tool does not resolve.
Who can call?
Cross‑origin, permissions and policies outside the model. The LLM chooses the tool; the site decides if it can run.
Where does the truth live?
Hybrid architecture: WebMCP on the page, authorization and rules on the backend, MCP for what stays outside the browser. No duplicate rules.
Is it really working?
Threat modeling, evals, logs, rollout and governance. That is what separates an experiment from a production operation.
05 · Tips for this week
Each tip corresponds to an item the scanner checks. Fix it, run again, see the score rise.
Note the overall score and the four sub‑scores. Blockers come first in the correction plan. It’s your baseline, and it’s free.
robots.txt saying what can be read, sitemap.xml updated and a llms.txt that presents the site to AI models. The scanner queries the three in every analysis.
Real title and description, one canonical per page, headings in order, images with alt, Open Graph and JSON‑LD describing the entity. None of this requires the Builder phase.
Real questions with direct answers, short definitions, lists and tables. That’s what the model can extract and cite. It’s the AEO score.
Entity name, authorship, update date and public evidence of what you claim. That’s the GEO score, and it’s what makes a model trust your content.
Take an existing form that doesn’t change anything dangerous, like a search or status query. It becomes your first declarative tool in the Builder phase.
06 · WebMCP Training from Zero to Expert
All open and in Portuguese. The phases use JavaScript and are for those who build or coordinate builders. A non‑programming site owner starts with the overview and diagnosis, and brings the phases to the team. Each phase has four modules of six topics.
Transform traditional sites into applications understood and operated by AI agents. Explains prerequisites, progression, how the four phases connect and the final project. Includes readiness diagnosis.
Open the overview →Transform sites into agent‑operable surfaces using declarative forms, JavaScript tools, JSON Schema, cancellation, fallback and a catalog validator.
Design coherent catalogs, model state and recovery, migrate existing sites and integrate WebMCP into modern applications without duplicating rules or removing the human journey.
Build browser agents that discover, select and execute WebMCP tools with validation, policies, permissions, cancellation and observable conversational loop. Authorization is outside the model.
Bring WebMCP solutions to production with threat modeling, hybrid WebMCP + backend + MCP architecture, evals, observability, rollout and governance.
WebMCP is still an experimental technology. The training makes the environment limitations visible and separates observable evidence from guarantees that depend on backend, authorization and human review. You leave knowing how to diagnose a site with critical sense, expose understandable tools, connect page, agent and backend without handing security decisions to the LLM, and assess risks, failures, observability and production criteria.
AEO and GEO playbook with the INEMA case. It is the half "search turned into ask" of this page, in depth.
Open the guide → The other side · the visiting agentFive flows for AI to operate the computer. Shows how the agent behaves on a site that still has no WebMCP.
Open the course → Business · 8 lessonsFor decision makers: what responsibility to hand to an agent, with scope, autonomy matrix and trust sheet. A site ready for agents is one of its tools.
Open the course →07 · Projects to get hands on
The scanner of this page. Independent project from the ecosystem, with its own repository, database and evolution cycle. JSON report with evidence.
Test the site → SpecificationSeven models that turned your agent specification on a page, with simulated trial. Used to design the agent that will consume your tools.
Open the guide → Agent with human gateTelegram gateway with durable queue, human approval and resumption. Layer 5 of this page, confirmation before acting, working on a real agent.
Open the guide → Step by step · 6 layersLayer‑by‑layer agent creation: Identity, Substrate, Rules, Skills, Tools and Agents. WebMCP tools enter the Tools layer.
Open the guide → Context and rulesHow a frontier model decides to use tools and ask for confirmation. Reference to write descriptions the agent understands.
Open the guide → In eventsA link turns into a video with human approval before heavy production. Example of chained skills with human gate, the same principle as the control layer.
See the page →08 · Start now
Scan it, see the evidence and join the training already knowing which decisions you need to master. Courses are open. What changes when you enter the ecosystem is having people, material and support instead of doing it alone.
The portal with all INEMA courses and projects: tracks organized by level, search by topic and the latest updates. It's where you start.
The community of those applying AI to real businesses: feed curated by Nei with what matters, groups by topic and support material from all trainings.
The platform for those who want to use AI to grow in practice: the more than 400 ecosystem projects, the trainings and the Brain with content organized for reference.
Unsure where to start? Talk to us on Telegram.