Methodology
How the registry is built, what it will and will not claim, and how to correct it. The whole thing is open data.
What gets listed
Agent Papers tracks the infrastructure that gives an AI agent the things a human has by default. We group it into seven categories: identity and trust, phone, email, address and mail, money, legal existence, insurance and liability.
The test for inclusion is simple. Does the company sell a product that gives an agent a form of personhood a human would otherwise have: an identity it can prove, a phone number, an email address, a mailing address, money it can hold and spend, a legal entity, or insurance against what it does. If the answer is yes, it belongs here. General developer tooling, model providers, and orchestration frameworks do not, even when agents use them, because they do not confer personhood.
A company can appear in more than one category when it genuinely ships in more than one. Each offering is listed on its own terms, with its own status and its own fields.
Blank over guess
A blank field is a real answer. It means one of two things: the company does not publish that detail, or we have not yet verified it from a source we trust. It never means the value is zero, and it is never a placeholder for something we assumed.
We do not estimate. If a company does not state its fees, the fees field stays blank rather than carrying a number we inferred from a pricing page that does not actually say it. Across the site, missing data reads as a quiet not published, so you can always tell the difference between a company that does not offer something and one that has not told anyone about it.
Sourcing
Every material fact carries a source. Each company page lists the URLs a claim was checked against, and those same sources travel with the data through the API and the registry file. The bar is a primary source: the company's own site, docs, or pricing page, or a filing, rather than a secondhand summary.
When sources disagree or a claim cannot be pinned to one, we leave the field blank and note what we know in the editorial section instead of picking a number.
Freshness
Every company records the date it was last checked. The dots you see around the site encode how recent that check is: a green dot means the check is recent, an amber dot means it is aging and due for a re-check, and a grey dot means it is stale. The pulse on the home page marks the registry as actively maintained, not any single company.
Freshness is about verification, not activity. A company can be very much alive while its entry ages; the dot is telling you how long since we last confirmed the details, so you know how much to trust them.
The graveyard
Dead and acquired companies stay in the registry. This is a young market, and what shut down or got bought is part of the map. Dead entries are shown with a struck-through name and kept in a separate section of each category so the market history is legible without cluttering the live comparison. Acquired companies point at whoever bought them.
Corrections
If something here is wrong, stale, or missing, tell us. Corrections are welcome from anyone, including the companies themselves. The submit page explains what a good correction includes and how to send it.
Open data
The registry behind this site is public and machine-readable. The same 60 companies you see here are available as:
- registry.json · the full dataset as one file
- a REST API · keyless, filterable, CORS-open
- an MCP server · for agents to query directly
- llms.txt · a map of the whole thing for language models