
Editorial score
It orders candidates for a particular query. It is neither a laboratory grade nor a success percentage nor an exhaustive comparison of models.
Our method prioritises practical usefulness, clarity and verification of public information. It explains why a tool may suit a task without turning a complex field into a definitive league table. Catalogue and official links checked on 27 August 2026.


We identify services with a clear proposition, recognisable public or professional access and a verifiable official website. A tool should make it possible to understand what it does, who it serves and how to begin. An announcement, a page with no identifiable access or a poorly documented experiment is generally not enough for the main catalogue.

Each tool receives a main category, dominant use, access mode and operational status. The publisher and country are included when publicly established. This does not attempt to summarise every feature; it supplies a simple landmark for finding the service and comparing it with nearby options.

Selection considers relevance to a concrete task, specialisation, apparent maturity, clarity of the offer and access to the official service. We also consider whether a tool has a useful place in its category. Popularity may indicate real adoption, but it does not replace an examination of features or fit.

Task routing uses words in the request, categories, declared uses and editorial scores. Results therefore change with the task. The routing does not measure a model’s absolute quality, guarantee an outcome or create a commercial ranking. Two very different tools may be equally relevant for different reasons.

Essential information is checked against available official pages, documentation, service access or retirement announcements. When a point remains uncertain, we avoid presenting it as established fact. Detailed prices are not copied because they change frequently and often depend on region, volume or account type.

Active, watch and retired statuses preserve context without presenting an unavailable service as operational. An update may correct an address, change a category, revise a status or remove an invalid direct link. Between reviews, the provider’s official page remains the most current source.

It orders candidates for a particular query. It is neither a laboratory grade nor a success percentage nor an exhaustive comparison of models.

Active marks an entry in the current catalogue; watch flags uncertainty or change; retired keeps a historical reference. Access describes the usual starting point, not the details of a contract.

Country refers to the organisation when known. It does not establish hosting location, subcontractors or data transfers. Those questions require the provider’s contractual documents.
To compare two research tools, ask the same precise question with a date and an expected source. Record the links supplied, open the original documents, check whether quoted passages support the answer and note the time spent correcting errors. A fluent answer without a verifiable source does not automatically perform better. Keep the question, date, results and checking criteria so the trial can be repeated after an update.
Prices, features, models, licences, usage limits and data policies may change without notice. A short entry cannot cover every contract, country or situation. Always check the official website before purchasing, deploying professionally or processing sensitive data. For important use, test quality on your own examples, review how data can be exported and keep a workable alternative.

The main use and functional category provide a first landmark without reducing the tool to a single feature.

Free, account, paid, local or enterprise indicates the usual way a person can begin using the service.

Active, watch or retired helps distinguish an available service from a reference retained for context.

The organisation and country are shown when publicly established, without assuming that this alone determines where data is processed.
The catalogue contains 317 references. 76 entries have a dated revision linked to an official source. An HTTP response confirms technical access; it does not validate every feature or contractual term of a service.
Country is recorded for 256 entries and left blank where the data does not establish it. A single country may not make sense for a community project. Country alone never establishes storage or processing location.
Former names remain identifiable to preserve links and favourites. Renamed products point to current references. Publisher-confirmed discontinuation is distinguished from network errors or anti-bot protections.
Detailed profiles and guides are available in French and English. A language in the selector does not mean all content is translated into it. Editorial illustrations depict usage scenarios, not providers’ real teams or premises.
Read profiles with sources and trials →Names and logos identify publishers and remain their respective trademarks. Generated editorial photographs are labelled as AI illustrations; they do not depict the tools’ real teams or premises. The logo additions below link to their sources; publisher marks are distinguished from tool-specific logos.
The evaluation workshop provides fictional documents, answer keys and explicit criteria. It covers the applicable document version, translation and JSON output, among other checks. Keep the answers and the model identity so you can repeat a trial.
Directory routing combines keyword matches, categories and editorial priorities. Suggestion order is not a performance ranking. A consultation date is shown on a profile only when recorded in its data; missing dates are disclosed.
Reproduce exercises with answer keys →
Choose an observable output, permitted data and an error requiring rejection. Family frameworks suggest checks to adapt before the first trial.

Check outputs against original documents. Record failures, corrections and human time alongside model version and configuration.

Success on a few examples does not guarantee all uses. Repeat checks after a change of model, documents, connector or access permissions.
Further reading on generative AI risk management and evaluation: NIST AI 600-1 — Generative Artificial Intelligence Profile (2024). This voluntary framework is a documentary reference; site examples and checklists are independent editorial proposals.
These figures are calculated from the displayed directory entries. They describe editorial coverage, not market share, popularity or tool quality. Services are counted in their primary family.
| Family | Active services | Profiles |
|---|---|---|
| Assistants | 19 | |
| AI search | 18 | |
| Image & design | 22 | |
| Video & motion | 14 | |
| Audio & voice | 12 | |
| Work & docs | 24 | |
| Code & build | 34 | |
| Local & open | 25 | |
| Agents & automation | 20 | |
| Cloud & enterprise | 18 | |
| Classic web | 3 |
For services marked active, these are the recorded outcomes of their latest source consultation. A readable page establishes neither feature availability nor service quality. These figures are not real-time checks.

Crawler protection, redirects and network errors require another check; they are not enough to establish discontinuation. Profiles link to their sources so you can inspect the context.
Knowledge bases, migration and vendor selection require different evidence. Explore the workflow that matches your decision.