The web was built for people who can see a page, find a button and fill a form. Agents cannot reliably do any of those three, and the gap is where a lot of this batch lives. Some tools make sites callable, some turn business APIs into tools an agent can invoke, some hand the agent facts it should never have guessed at, and one is a reminder that a great deal of agent output is still unverifiable prose.
Making a website callable
Latch is the most direct answer: one script tag that makes a site work inside AI browsers. Agents already click through storefronts and fail on buttons and forms, so Latch exposes search, cart and contact actions as WebMCP tools. It registers search_site to find and submit your search box, add_to_cart to click buy controls, submit_form to map your fields, and navigate to surface your main nav. It reads the page read-only, is about 6 KB with zero dependencies, needs no server and no rewrite, and degrades silently where WebMCP is not present.
The caveat is real: WebMCP is a W3C standard currently in a Chrome origin trial, with Edge, Atlas and Comet listed. That is early. But the bet is cheap, and being callable when agent traffic arrives is better than being unclickable.
Business APIs as tools
Rivinci runs both halves of a business agent: the chatbot that talks to people and the MCP servers that turn the APIs you already have into callable tools. Put together, the agent does not just reply, it acts. It answers customers on your website, WhatsApp and Telegram, and Rivinci Cloud is in public beta with a free starting tier.
APIStack Facts MCP solves the narrower and more common problem of agents confabulating facts. It offers sixteen deterministic tools for the things a model should not answer from memory: the current time in a time zone, holidays and working days, exchange rates, VAT and IBAN checks, company lookups, link safety and sanctions screening. You add the URL and it works with no sign-up, with fifty free calls a day and an optional free key beyond that.
Facts a machine can check
On the hardware side, Quantized VRAM Index answers the question every local-model user eventually asks: what actually fits my GPU. You pick a card, from an RTX 5090 with 32 GB down to older Tesla and Mac configurations, and see which of 123 models fit, measured on real hardware rather than estimated. It also carries a quant hub, benchmarks and a cookbook.
Workflows as text
Requ is an open-source diagram language for people and LLMs. You write a workflow or operating map as plain text and see it render as you type, defining teams, customer journeys, AI agents, handoffs and swimlanes. Because the source is text it works in the playground, on the command line and in version control, and it exports SVG, Markdown or an AI-ready prompt. It is MIT licensed at version 0.1.
Orchestration and delegation
Seclai is the broadest entry here, positioning itself as everything around your LLM calls: workflow orchestration, output evaluation, governance and switching between 200+ models, visually or in code, with a dashboard covering traces, error rates, knowledge bases, memory banks and human-in-the-loop oversight.
Kavela sits at the other extreme. You write a task in plain words and it runs it in your tools and checks every run. The published example reads the Hacker News front page, judges relevance, picks the three most relevant AI-agent posts, writes a five-line summary to a file and reports back. The promise in the name is the useful part: automations that do not fail silently.
A necessary note of scepticism
Kouh AI Text Detector rounds out the batch. It scores writing for AI authorship entirely in the browser, highlighting AI-like passages and giving an overall percentage with a confidence level, with nothing uploaded or saved. Notably it states plainly that no detector is perfect and the score is an indication rather than proof, and it needs at least fifty words before it will say anything at all.
The awkward middle
What all eight are really working on is the seam between an agent that can reason and a world that was not built to be reasoned about. A storefront expects a human to see a button; a business API expects a client that knows its schema; a date, a holiday and an exchange rate all expect someone to look them up rather than recall them. Each tool here is a small adapter across that seam, and the good ones are explicit about what they do not cover.
That is where the scepticism belongs. A detector that admits it is fallible, a facts server that only answers what it can verify, an index measured on real hardware rather than estimated: these are the entries that will still be useful in a year. The ones that promise to handle the seam invisibly are the ones worth watching closely before you depend on them.
In a batch full of tools reaching further into the world, that honesty is worth keeping in view.
