A reference implementation, not a product
AgentDossier shows what can be found about an AI agent before anyone signs a contract: who publishes it, which certifications the registries and the vendor's own pages name, which protocols it serves at its well-known endpoints, and what has been written about it lately. It is a working demonstration of an evidence-first agent registry, built and maintained by one person, and the code is open so that the method can be checked and reused. It is not an assessment, a certification or a recommendation.
Who
Built and maintained by Hemant Naik
(hemant.naik@gmail.com). There is no company behind it and no vendor
pays to be listed. Curated entries that carry a checked_by name were checked by that person on the date shown.
How the data is produced
| Step | Automated or by hand |
|---|---|
| The Top 100 lists per domain | A curated workbook (last compiled 2026-09-25) chooses which agents appear. Their order is recomputed on every build: trust and governance come from the evidence; adoption, health, ecosystem, fit and docs are the workbook's values. |
| Discovered agents (823 records in total as of 2026-10-26) | Automated: GitHub, Hugging Face, the MCP registry, marketplaces and ARD manifests, with demo, deprecated and inactive projects dropped. |
| Compliance evidence | Automated matching against the FedRAMP Marketplace and CSA STAR, plus a crawl of each vendor's trust pages (claims, tier 4). A small curated file holds records a person verified. |
| Publisher identity | Automated rules (ARD manifest, marketplace listing, vendor-domain page, repository ownership) plus a curated file of domains a person confirmed. |
| Protocols (MCP · A2A · ARD) | Automated probes of each publisher's well-known endpoints; curated claims from vendor documentation carry a reviewer's name. |
| Worked comparisons and the reference set | Answers are collected from vendor documentation with AI assistance; a comparison is marked draft until the maintainer records a review. |
| Traces and the incident registry | Curated: one row per real, sourced event where an AI agent attacked, defended or was the target, "as reported", each row saying who reviewed it. A weekly scout reads public feeds for candidates; a weekly machine review reads each candidate's source and publishes it when it describes a real event (marked machine-reviewed, decision on record) or dismisses it with a reason; a weekly re-check records whether every source link still answers. Per agent: the registry's own endpoint probes and vendor-published bundles in a hash-chained store. The harness produces tier-1 traces only once a model key is configured. |
| The CVE registry | Automated from the official record: every CVE (via NVD) that names a listed agent product or whose record is about agents (CWE-1427 or the agent vocabulary), with CISA KEV joined. Official severity and descriptions; this registry's category, surface tags and cross-links; a person's overrides on file. Never a score, never an exploit. |
| News and chatter | Automated: Hacker News, GDELT, GitHub releases, vendor feeds and NVD. News never changes a score. |
How often
The catalog is rebuilt every Monday by a scheduled job — the same run scouts public feeds for candidate incidents, reads and publishes them with a model, and re-checks every incident's source links — and the date of the last build is shown in the banner on every page. Each build appends to the trust changelog, so a badge that appears, changes or disappears leaves a record.
Reporting an error
Every dossier has a Dispute or correct this dossier link that opens a pre-filled GitHub issue. Corrections with a source are applied in the next build; curated files record who checked what and when. The methodology page explains every tier, score and status on the site.
What this site does not do
- It does not verify certificates with issuers. A registry match means the registry lists it; a vendor claim means the vendor says so.
- It does not rank agents by safety. Scores are discovery signals under a published formula, and "unknown" never means non-compliant.
- It does not track visitors. No accounts, no cookies, no ads; the public demo sends only aggregate, cookieless page counts.
Code: github.com/Sushegaad/AgentDossier (MIT) · catalog data CC BY 4.0 · self-hosted enterprise edition for private catalogs.