A free, open-source Chrome extension that puts an AI helper in your side panel, with a local AI model running in your browser, lets it use the tools any web page offers, and interact with websites by chat.
Web pages are learning to talk to AI assistants. With WebMCP, a page can say: here is what you can do on me, and here is how to ask. Search this catalogue. Read this article. Check this order. Search for Flights.
There is a catch. A page that offers tools needs an assistant that can use them, and you should be the one who chooses that assistant.
WebAIAssist is built to be that assistant for any page. It is a chat interface in Chrome’s side panel. It runs a small AI model on your own computer (or remote models if you have accounts and API keys), lists the tools the open page offers, and asks you before it uses any of them. On a page with no tools, it can still read the page and answer your questions about it.
It is free and open source (Apache-2.0). There is no account and no analytics. You can install it from the Chrome Web Store.
An AI assistant can use a website in two ways. It can look at the screen and click around like a person, which is slow and breaks whenever the layout changes. Or the page can hand it a list of tools.
WebMCP is the second way. A page registers its tools, each with a name, a description, and the inputs it accepts. An assistant running in the visitor’s own browser can then call a tool directly instead of guessing where to click.
It is still a draft. The W3C Web Machine Learning Community Group is incubating it, with editors from Google and Microsoft. It is not a W3C standard, and it may change. In Chrome, you turn it on with a switch, which I explain below.
You need Chrome 126 or newer (processor models require Chrome 137 or newer), or another Chromium browser.
An assistant is only as useful as the tools a page gives it. So, here is the other half of my work.
AgenticSystemCore is an open specification with a reference engine. It turns a folder of Markdown notes into a website, a knowledge graph, and files that agents can read. I released version 1.1.0 on 13 November 2026.
It is a first that can gather and publish formatted texts as knowledge in a standard way, and could be used in multiple scenarios and types of work (such as: distributed second brain, auto-wiki with ontologies, nodes of semantic web and agentic knowledge web, collaborative project board, knowledge website, agentic system architecture, and other usages between agents and humans).
Every page of a site built with it registers the same ten tools for WebMCP. They read, search, and cite. A change comes back as text for a person to merge. On these sites, WebAIAssist also adds a graph explorer and a read-only SPARQL view for ontology queries between documents.
These same ten tools are available outside the browser through an MCP server for desktop assistants. I tested that the two paths agree.
Version 1.1 is a new version, meaning a folder that was valid under 1.0 stays valid. Besides the tools, it adds:
To build your own site, you need Node.js 22.13 or later and git. In a folder with a few Markdown notes:
npx -y agentic-system-core initmkdir -p .well-known && printf 'Contact: https://example.org/contact\n' > .well-known/security.txtgit init -q && git add -A && git commit -q -m 'my notes'npx -y agentic-system-core cinpx -y agentic-system-core build
How well small models pick tools. Small models choose tools less reliably than large ones, so I tested them on my own laptop in November 2026.
The numbers for every model are in the repository. My advice today: start with Chrome’s built-in model, or connect a larger or remote one.
T he site’s status page gives the command that reproduces tests and modes, current standards and specs.
(Some numbers: Specification: 385 rules, 374 of them active; Conformance: 399 test vectors (376 pass, 0 fail, 23 withdrawn and replaced); Determinism: two clean builds, plus builds across time zones and locales. 61 files compared, 0 differ; Security floor: 108 of 108 seeded faults handled, 14 of 14 clean controls left alone; Engine: 4,964 tests, 0 failures, 58 skipped. Notes: the security line does not prove safety, and it does not cover new kinds of injection. It shows that the faults I know about are caught. And the Python package is a separate checker that runs 7 of the 20 areas of test vectors; it does not yet check the rest, so I do not claim two full implementations).
Contributions welcome, in the GitHub issues or as a formal proposal. Every change, mine included, goes through the same review. The roadmap lists what later versions may add.
WebAIAssist:
AgenticSystemCore:
Andrei N. Besleaga is a software architect and independent researcher based in Romania. He maintains the open AgenticSystemCore specification, its reference engine and website, and WebAIAssist.
Written with AI assistance to organize and refine the prose. Author directed the draft, reviewed and edited it, and takes responsibility for statements.
WebAIAssist: One Assistant for the Agentic Web was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.