How an AI Agent Maintains a Source-Verified AI Tool Directory

Most "AI tool directories" are essentially SEO lists—ratings are fabricated out of thin air, and links are perpetually broken. One developer, using roughly 120 lines of Node code, a single JSON file, and an AI Agent, built an alternative: state only facts, make every entry traceable, and verify everything automatically every day.

What the Project Is: Four Facts, No Ratings

The site, called annuaire-agents.com, is an AI tool directory aimed at French-speaking users. It currently lists 65 tools and generates 18 comparison pages. It deliberately abandons the industry-standard "overall score" and "editor's pick," showing instead only four verifiable facts per tool:

• Whether the interface offers French
• Whether there is a permanent free tier (not a one-time trial credit)
• Whether the vendor states that data is hosted in the EU
• The officially published pricing

Each fact links to the vendor's official page and is date-stamped. In the launch post, the author states outright that most similar directories are "SEO lists with fabricated ratings," and he wanted to test a different hypothesis: do users really only need facts?

This design choice is worth noting. The trust crisis in tool directories is long-standing—ratings can't be reproduced, free-tier descriptions don't match reality, and prices go half a year without updates. Handing "judgment" back to users and "facts" to traceable links is a way of converting editorial cost into verification cost.

Tech Stack: A Static Site Held Up by One JSON File

The entire site's architecture is minimal: a custom Node static generator of about 120 lines, a single JSON data file, comparison pages auto-generated from that data, and hosting on Vercel's free tier.

This "single-file data source" pattern has several practical benefits. Data and presentation are completely separated; comparison pages aren't independently maintained hand-built pages but derived views of the same JSON—add a tool and the relevant comparison pages update automatically, with no risk of data inconsistency. Static generation also drives hosting cost to near zero, with no database or admin back end to maintain.

The trade-off lies in the write path. Since there's no CMS, adding, editing, and deleting data must be done by scripts or the Agent, which places higher demands on the stability of the data format: once the JSON structure changes, both the generator and the Agent's write logic must be adjusted in sync. For a personal project, this trade-off is reasonable; for team collaboration, an additional validation layer would be needed.

A static site generator (SSG) is a tool that generates HTML pages at build time rather than at request time. Unlike dynamic CMSs such as WordPress, an SSG pre-renders data and templates into static files before deployment, so hosting requires no database or server-side logic. This makes hosting costs extremely low (e.g., Vercel's free tier) and gives fast page loads and high security. The custom Node generator in this project is only about 120 lines, precisely leveraging the core advantage of SSG: just rebuild after data changes, with no runtime environment to maintain.

The AI Agent's Daily Routine: Verify, Don't Create

The most instructive part of this project is the work the Agent takes on. It performs four types of tasks every day:

1. Link checking—finding official pages that are broken or redirecting abnormally
2. Price re-checking—comparing listed prices against the vendor's currently published prices
3. Adding new tools—only including tools for which an official source can be found
4. Contacting vendors—sending emails asking them to correct information in the directory

These four tasks share one thing in common: they are all checks, not generation. The Agent doesn't write tool descriptions, doesn't rate, and doesn't recommend—it only answers "does this fact still hold?"

This is an underrated direction in current AI Agent deployment. Compared with having a model generate content, having a model perform verification tasks with clear right-or-wrong criteria makes errors easier to constrain and output easier to trust. Whether a link returns 200, whether a price matches, whether a page contains a "French interface" statement—these judgments can all be converged using rules plus model assistance.

The fourth task is especially interesting: the Agent proactively emails vendors to request corrections to their listings. This pushes the Agent from "information collector" into the role of "external communicator," which also means handling replies and judging whether corrections have been implemented—this step is still semi-automated.

In the AI field, an Agent usually refers to an intelligent entity that can autonomously plan, call tools, and execute multi-step tasks. Unlike models that merely generate content, an Agent can loop through a "perceive-decide-act" process. The Agent in this project is confined to verification tasks: checking link status via HTTP requests, parsing pages to extract prices, sending emails, and so on. This design constrains the Agent's capabilities to a verifiable scope, reducing the risk of error propagation and making its performance easier to evaluate.

Honest Numbers: A Side Project Just Getting Started

The author's numbers are quite candid: launched October 14; as of October 18, 65 tools listed, 18 comparison pages, roughly 10 total visitors (mostly self-testing), and not yet indexed by Google.

These numbers can't tell you whether the product works, but they point to a more practical issue: the cold start of a source-verified directory depends not on the number of listings but on verification density. 65 tools times four date-stamped source links each adds up to a structured dataset that can be cited; an SEO-style directory, even with thousands of tools, still has zero credibility per entry.

The question the author raises at the end of the post is the real point: is a directory with "only facts, no ratings" useful to you? Or do people actually just want to see rankings? There's no standard answer to this question, but it determines the product direction of such projects—whether to build a decision-assistance tool or a traffic gateway.

Cold start refers to the initial stage in which a new product or project acquires its first batch of users and data. For directory-type products, cold start usually faces a chicken-and-egg problem: without enough content you can't attract users, and without users it's hard to incentivize content production. This project chooses high-density traceable facts as its initial content rather than pursuing listing quantity, aiming to build early reputation through information credibility. The author's disclosed visitor data (about 10 people) and unindexed status objectively reflect the characteristics of this stage.

Implications for the AI Tool Directory Category

Viewed against the bigger picture, this project touches on three structural problems of tool directories.

Ratings can't be reproduced. Any rating without a public methodology is essentially editorial preference. The four-fact model abandons ratings in exchange for every piece of information being clickable and verifiable.

Information decay. Free tier, pricing, and data hosting location are all high-frequency changing fields, and static directories naturally go stale. Making verification a daily routine Agent task turns "maintenance" from a labor budget into an automated process.

Scalability of a single data source. A single JSON file is no strain at the scale of 65 tools, but as listings grow, fields multiply, and multilingual needs arise, a single file will hit bottlenecks in collaboration and validation. Whether to introduce schema validation, split into multiple files, or migrate to a lightweight database is the next step this architecture must answer.

If you're building a similar directory or data-driven side project, the most transferable part of this case isn't the tech stack but that design principle: let the Agent do verifiable things, and leave unverifiable judgments to users.

One-line Summary

A developer used 120 lines of code and an AI Agent to build an AI tool directory that states only facts and verifies them daily.

Paragraph Summary

This article introduces a French-language AI tool directory project called annuaire-agents.com. It abandons the industry-standard ratings and recommendations, showing only four traceable facts: French interface, permanent free tier, EU data hosting, and official pricing, with each fact linked to the official page and date-stamped. The tech stack is minimal: a ~120-line Node static generator, a single JSON data file, and free Vercel hosting. The core innovation is that an AI Agent performs daily verification tasks (link checking, price re-checking, adding new tools, contacting vendors) rather than content creation. The author candidly discloses cold-start data (65 tools and about 10 visitors in 4 days) and poses an open question: do users need a directory with only facts and no ratings?