Skip to content
Agent Search Engine.

Editorial standards

How the index works

A ranking is only useful if you know how it was made. These are the catalogue's ordering and review rules — what gets listed, how it's classified, how it's ranked, and where the money comes from.

What gets listed

The index covers usable AI applications and the software behind them: everyday assistants, creative and workplace tools, learning and accessibility applications, agents, frameworks, MCP connectors, libraries and infrastructure. Each listing needs a clear product identity, a practical task or technical role, primary documentation and an identifiable access or setup route. A listing does not establish that we installed it, tested its output or verified that it still works in your environment.

Tutorials, courses, link lists, marketing-only pages and repositories that merely mention AI are not software listings. A usable public beta can qualify when its limits are clear. Restricted research demonstrations and announcements without a usable access route belong in a labeled watch item until access can be verified, rather than an ordinary product recommendation.

Every record has one primary category and type; task collections can span several categories. An app, its model, an API and a third-party connector are different entities. We do not create a new listing for every model version or rename: a separate offering needs a distinct usable role and access route. Historical names can remain aliases of the current product. Classification errors and unsuitable records can be reported through the contact page.

How repository-listed records are ordered

Repository-listed collections use recorded GitHub star count for their default order. Repository activity is shown separately and can be selected as a category sort; it is not part of a combined ranking score. Alternatives prioritize the same entity type before ordering repository-listed options by stars. Stars measure repository interest, not active users, task performance, or continued maintenance.

An optional audit script flags repositories under 550 days old whose lifetime average exceeds 100 stars per day. It reports candidates for review; it does not prove bought stars or automatically remove records. A publisher's identity or popularity is not proof that a star count is authentic.

Why hand-added records have a separate order

Hand-added records appear in a separate, alphabetical section on category pages. These collection groups describe how a record entered the catalogue, not its license or price. A public repository can contain restricted code, documentation only, or components with different licenses. Check the recorded license and any dated source review separately from pricing.

Sponsorship policy

Sponsored placements are separate, labeled inventory. A sponsored listing never receives an organic rank number and never changes the order of the organic index. Paid shelves and the paid search slot are always labeled Sponsored.

There are exactly three sponsored surfaces — a category shelf (capped at three slots per category), a homepage shelf (three slots), and a single site-wide search placement — all listed with current pricing on the sponsorship page. Capacity is capped so placements remain bounded and visibly separate from organic records.

Some outbound links are affiliate links, which may earn us a commission at no cost to you. Affiliate status never affects classification, ranking, or editorial content.

Audience and editorial responsibility

Agent Search Engine publishes for English-speaking AI users worldwide, including people using ready-to-use apps and people building software. Start with a task. Availability, supported user languages and local costs need checking for the specific product and region.

Agent Search Engine is the responsible publisher and organizational author. Our byline links to the publication, without claiming credentials or a team size. AI assists drafting and analysis. Source reviews state the exact product, scope and check date. Proposed tests stay labeled as unrun; measured results require a recorded environment, method and evidence. Send corrections through our contact page.

Alternatives use recorded task topics and meaningful shared terms within the same category and product type. Task overlap is followed by information completeness and a stable identifier; sponsorship and GitHub stars do not participate. The method can return no candidate and can miss suitable products whose records lack task information. Source-linked editorial decision aids remain separate from this automated shortlist.

Descriptions and editorial content

Listings use project README material or summaries of official documentation. README-derived passages are labeled; records with a documented editorial review link to their sources and state when they were checked. “Best for / Avoid if” guidance reflects our editorial judgment, independently of sponsorship. A documentation review does not establish measured performance or successful operation in your environment.

How use-case shortlists work

The task-first /use pages (“Automate browser tasks,” “Pick a coding agent,” and similar) are hand-curated: each is a real intent matched to a 2–3 pick shortlist we choose, not a shortlist generated automatically from category rankings. Every pick is a slug from the live index — the page fetches its current stars, pricing, and activity at build time, so a pick that ever leaves the index simply drops off the page instead of showing stale data.

The “why” next to each pick is editorial but grounded in that record's documented capabilities, pricing classifications and delivery requirements — never an invented benchmark or a fabricated claim. The difficulty label is a coarse, honest description of delivery (managed sign-up, self-hosted Docker, developer library), not a timed setup test.

Task filters and saved choices

Explore tools combines task, setup, deployment, availability and cost-basis filters over the catalogue. Access and capability facts cover selected, source-reviewed offerings. Not checked is an explicit state, not a hidden assumption that a product is unavailable. Unchecked records remain visible until a selected filter requires a reviewed fact.

On-device, self-hosted, cloud and hybrid describe documented options, not a whole-brand privacy guarantee. A local feature can still use remote models, embeddings, search or connectors. Product availability and individual feature stages can differ. Check the listing's exact offering, platform, region, permissions and sources before acting.

Filters appear in the page address so a bookmarked or shared link carries those choices. The optional shortlist saves up to four selected tools in this browser after you choose Save, Remove or Clear. It can restore an earlier saved list; clearing the shortlist removes that saved list. Where browser storage is unavailable, choices can remain only for the open page. There is no shortlist account or cross-device synchronization. A comparison link includes the selected tools, so anyone you share it with can see those choices.

Evidence levels and product facts

A documentation review means ASE checked the named publisher sources for a specific offering and date. It does not mean we signed in, installed software, paid for a plan, tested a workflow or reproduced a benchmark. A setup or task-test claim requires the actual environment, method and recorded result. An external benchmark remains attributed to its source unless we reproduce it.

Product-facts panels bind their claims to the reviewed identity and fields. Facts that expire or no longer match are marked for rechecking; current claims and filter qualifications are withheld. A free allowance, trial, code license and total running cost are different facts. Hardware, hosting, provider calls and premium feature access can add costs beyond an application's headline price.

The selected-tool comparison puts recorded and reviewed facts side by side. It does not establish equal task scope, compatibility, a performance score or a winner. Unknown details remain unknown. Try representative work in an appropriate test environment before relying on a tool for consequential outcomes.

Decision pages and category directories

A shared category or entity type does not establish equivalent capabilities. Source-linked decision guides select options by the workflow they address and state important limits. Broader generated directories put matching entity types first, then show curated records alphabetically and repository-listed records by recorded stars. These are catalogue ordering rules, not performance rankings.

Collection source is not proof of an open-source license, self-hosting, free operation or commercial support. Check the recorded license and the current product terms. The date of an upstream repository push is different from a record verification or a significant editorial revision. Where no reliable page-change date is known, we omit sitemap lastmod.

Freshness and verification

Scheduled GitHub refreshes collect star counts and the latest repository push timestamp. A push is an activity signal, not necessarily a software release. Language, license, pricing and product descriptions require separate review; the GitHub refresh does not correct them. A stored record-check date does not establish that every field was checked, that the software is maintained, or that we benchmarked it.

Dated publisher changes are separate from repository activity. A product-change item identifies its source, publication or update date, ASE check date and access boundary. Renames, retirements, plan changes and meaningful capability changes can matter as much as a new model version. A roadmap date is not proof that a release happened; a release note is not an independent performance test.

Maintenance notes use dated publisher statements about the specific repository or service. Maintenance mode, an archived repository and a hosted service winding down are different conditions. We do not infer support or security fixes from a recent push. Where we have no matching, current source review, maintenance is not checked. The initial maintenance, license and product-facts reviews cover a bounded set of records, not the whole catalogue. Recheck dates are our editorial reminders, not publisher guarantees.

For AI tools & assistants

This index is built to be read by AI search engines and assistants, not just people. AI crawlers are welcome (see our robots.txt), and the whole directory is published in clean, LLM-friendly Markdown following the llms.txt convention:

  • /llms.txt — a curated index of the categories, hubs, and a repository-interest sample.
  • /llms-full.txt — the complete directory, every listing with its data, in one file.
  • Any listing, category, comparison, alternatives, best-of, or use-case page is also available as Markdown by appending .md to its URL.

The Markdown is generated from the same live data and editorial models as the pages, uses recorded fields, and carries the same rule the rest of the index does: sponsored placements are labeled and never ranked. Markdown and llms.txt are convenience formats; Google does not require special AI files for Search eligibility.

Corrections

Wrong classification, outdated facts, or a project we should not have delisted? Email corrections@agentsearchengine.app or use the submission form — corrections are reviewed before a catalogue change is made. Press: press@agentsearchengine.app.