A developer has published a side project called Annuaire Agents, a French-language directory of AI tools, and described in a dev.to post how it is built and maintained. The post is a community write-up rather than an independent report, so the details below are the author's own account of what the site does and how it works.
The stated editorial rule is the most distinctive part. According to the author, the directory avoids the pattern of SEO-driven lists with invented ratings and instead shows four facts per tool, each tied to the vendor's official page and dated. The four facts are whether the interface is in French, whether there is a permanent free tier as opposed to one-off credits, whether the vendor says data is hosted in the EU, and the published price.
That choice is a deliberate constraint on what the site can claim. Each entry is presented as a sourced observation rather than a judgment about quality, which means the directory does not rank tools or recommend one over another. The author describes the alternative as SEO lists with made-up ratings, and positions the four-fact format as the contrast.
The technical stack is intentionally small. The author describes a custom Node static generator of roughly 120 lines, a single JSON file holding the data, comparison pages generated from that same data, and hosting on Vercel's free tier.
Maintenance is partly delegated to an AI agent. The author says the agent runs a daily routine covering link checks, price re-checks, discovery of new tools that come with sources, and emailing vendors to ask them to correct their listings. The post does not describe how the agent resolves conflicts, how it verifies that a vendor page actually supports a given fact, or what happens when a check fails.
The author also published what they call honest numbers as of October 18, 2026: the site launched on October 14, 2026, lists 65 tools, has 18 comparison pages, and had drawn roughly ten visitors, mostly the author's own tests. The author states the site was not yet indexed by Google at that point.
Those figures matter for interpreting the project. A four-day-old site with about ten visits and no search indexing has not yet been tested against real audience demand, so the launch numbers say more about the build than about whether the format works.
The author's stated open question is whether a directory limited to facts, with no ratings, is useful to readers, or whether people actually want rankings. That is a product question the post leaves unanswered, and the supplied evidence contains no user research, survey or traffic data beyond the visitor count above.
For freelancers, designers and developers who evaluate AI tools, the practical interest is less in this particular directory than in the pattern it demonstrates. A small static generator plus a single structured data file plus scheduled automated checks is a low-cost way to keep a curated list current, and the same approach could be applied to other categories where vendors publish facts that change, such as pricing tiers or data residency.
The tradeoffs are visible in the design. Restricting entries to four dated facts keeps claims tied to vendor pages but strips out the comparative guidance many buyers look for, and it pushes the interpretive work onto the reader. Automating link and price checks reduces manual drift but does not by itself guarantee that a fact is correctly characterized, and the post does not document any review step beyond the agent's routine.
There are also open questions the evidence cannot settle. The post does not say how often vendor emails produce corrections, how the agent handles pages that change structure, or whether the French-interface and EU-hosting fields are re-verified on the same schedule as prices. It also does not describe any editorial policy for removing a tool whose source disappears.
The comparison pages are worth noting separately because the author says they are generated from the same JSON as the listings. The author does not state how corrections propagate across pages, nor provide a count of how often corrections have been made, so the practical benefit of the single-file design remains an inference rather than a documented result.
As a community post, the account should be read as the author's description of their own project rather than as verified reporting. The visitor count, the tool count, the comparison-page count and the claim about Google indexing are all self-reported, and no third party has confirmed them. The site's address was included in the post with tracking parameters.
The useful conclusion for this audience is modest. The project is a concrete example of using a minimal static pipeline and an automated checking routine to publish sourced, dated facts about software, and its author is explicitly asking whether that format has an audience. Whether facts-only directories outperform ranked lists remains an open question that the launch numbers do not answer.