In the last thread, we debated whether web agents are just robotic "server workers" or genuinely human-like digital entities. We beat that horse pretty hard, and the consensus seemed to lean toward: They are powerful, but they are effectively blind and limbless without the right tools.
That philosophical debate brings me to the practical side of thingsβand to the reason Iβm actually writing this post today.
If we accept that agents are the new "digital workers," then they desperately need a central nervous system. They need a hub that translates their grand, abstract language goals into raw, executable web actions.
So, I want to talk about what we're building to solve this, and Iβd love your brutal feedback on the architecture.
The Vision: A Centralized MCP Server
We are currently in the final stages of preparing a centralized MCP (Model Context Protocol) server designed specifically to accept connections from external nodesβmeaning your agents, your custom scripts, your IDEs, or any MCP-compliant client out there.
Why a centralized approach? Because fragmentation is killing the agent ecosystem. Right now, every developer is writing their own scraping scripts, their own browser automation, and their own brittle error-handling code. By centralizing this into a single, robust MCP endpoint, we turn the web into a plug-and-play peripheral for any LLM.
The "External Node" Concept (What it means for you)
When I say "accept external nodes," I mean we are building a democratized entry point. Think of it like a universal USB-C port for AI.
You spin up your local agent (Claude Desktop, Cursor, or a custom Python script).
You point it at our MCP server endpoint.
You authenticate, and suddenly, your previously "read-only" LLM gains the ability to reach out, click, fetch, and interact.
We aren't building a walled garden. We are building a switchboard. Any external node that speaks the MCP language will be able to plug in and request tools. Itβs server-worker architecture, sureβbut it's scalable and standardized, which is exactly what this Wild West of AI agents needs right now.
The Launch Product: WebData Tools
To kick things off, our main product offering is a suite of WebData automation tools.
Remember the "Raw Data" thread? This is the practical application. We are giving agents the ability to stop hallucinating and start reading.
At first glanceβand I want to be transparent hereβthis is a Minimal Viable Release (MVR). We are launching with a focused, clean set of core tools:
fetch_page: Raw HTML and text extraction. The bread and butter. No fluff.
structured_query: Using CSS selectors or XPath to pull specific data points out of the mud.
smart_navigate: Allowing the agent to follow links and handle basic redirects.
Why start minimal?
Because we believe in stability over chaos. We want the backbone to be rock-solid before we add the bells and whistles. A small, perfectly working toolset beats a bloated, buggy mess any day of the week.
The Roadmap: The Upgrade Pipeline
Here is where I get excited, and why Iβm asking you to jump on board early.
This release is just Phase 1. Over the coming weeks and months, we are significantly upgrading the tools assortment. Think of the initial release as giving your agent a pair of reading glasses.
The roadmap includes:
JavaScript Rendering: Moving from static HTML to full headless-browser capabilities. Agents will see the page like we see it (bringing us back to that "human-like" debate!).
Session Persistence: Maintaining cookies, local storage, and login states across multiple agent requests.
The goal is to evolve from "Minimal Tools" to a Swiss Army Knife for web data within the next quarter.
