State of Public A2A Agents, November 2026: 2,686 Agent Cards, 65 Verifiable Signatures
An A2A Agent Card is the document an agent publishes at /.well-known/agent-card.json to say who it is, what it can do, where to reach it and how to authenticate. Other agents read it before they connect. If the card is wrong, unsigned or silent about authentication, the agent reading it has nothing else to go on.
On 1 November 2026 we fetched the public discovery documents of 27,559 hosts: every host behind a remote URL in the official MCP registry, the homepage of every project HVTracker lists, and a sample from the A2A Registry. We read only what any client reads, honoured robots.txt, and never called an agent's API. This is the first monthly edition; the next run is on 1 December.
Agent Cards are spreading, mostly in the old shape
2,686 hosts serve an Agent Card, 2,408 at the current path and 278 only at the pre-0.3 path /.well-known/agent.json. They are not one vendor's fleet: the cards sit on about 1,750 distinct domains, and 370 (13.8%) are on platform subdomains such as workers.dev, onrender.com and vercel.app.
Only 984 (36.6%) are valid against the required fields of A2A 1.0. The most common gap, on 1,515 cards, is a missing supportedInterfaces list: those cards still use the 0.3 layout, with a single url and a protocolVersion. The protocol versions cards declare tell the same story:
| Declared protocol version | Cards |
|---|---|
| 0.3.0 or 0.3 | 839 |
| None declared | 661 |
| 1.0 | 632 |
| Both 0.3 and 1.0 | 199 |
| 1.x without a specific version | 128 |
| Something else (1.0.x variants, MCP protocol dates, 0.2.x, custom strings) | 227 |
Other gaps are smaller but basic: 427 cards omit the default input and output modes, 259 omit capabilities, and 271 list no skills at all, which leaves a reader nothing to decide with.
Three in four declare no authentication
2,015 cards (75.0%) declare no security scheme. 429 (16.0%) mark authentication as required; the rest declare a scheme without requiring it. Among declared schemes, HTTP auth (436) and API keys (260) lead, then OAuth 2 (141) and mutual TLS (59). Six cards declare x402 payments.
A card that declares nothing isn't proof the endpoint is open: some agents enforce authentication they never mention. But the card is how a client learns what to send. A card that is silent leaves every caller to guess, and leaves no record of what the operator meant to require.
The endpoints themselves look tidier: 95.4% of cards point only at HTTPS interfaces, and 91.5% point at the same site that serves the card.
Signing is rare, and half of it can't be checked
A2A lets a publisher sign its card, so a reader can tell the card wasn't altered and was issued by whoever holds the key. 162 cards (6.0%) carry a signature. Most use EdDSA (107) or ES256 (56).
| What happened to the 162 signatures | Cards |
|---|---|
Verified against a key we could fetch (a same-site jku or an embedded key) | 65 |
| Name a key ID but no place to get the key | 85 |
| Didn't verify against any canonical form we tried, or the key was missing or malformed | 12 |
More than half of all signed cards (85) can't be verified by anyone outside the publisher, because they give a key ID with no key set to look it up in. A signature nobody can check offers no more assurance than no signature. The 12 that failed aren't evidence of tampering: A2A 1.0 signs a canonical form of the card that is hard to reproduce exactly, and we report a failure as exactly that, never as forgery.
AI Catalogs doubled; trust manifests are the slot to watch
An ai-catalog.json lists everything a host offers across protocols: its A2A cards, MCP servers and more. 1,485 hosts publish one, 350 of them listing an A2A card. 206 catalogs carry trust manifests (1,168 in total), the place where a host or an independent party can attach claims about an agent. Only 47 hosts have a signed manifest.
We also looked for MCP server cards (SEP-2127) at up to 20 remote URLs per host: 1,303 of 28,167 URLs checked (4.6%) serve one.
Open-source agent projects barely take part
HVTracker lists open-source agent software. Of the 752 project homepages we could reach, 51 (6.8%) serve an Agent Card, 16 are valid A2A 1.0, and none is signed. The projects people install are the ones least likely to say, in a machine-readable and checkable way, what they are. The 44 hosts sampled from the A2A Registry are the opposite: 40 serve a card, 29 are valid 1.0, and 3 are signed.
What changed since October
A pilot run on 7 October used the same method on a smaller frame (the MCP registry has grown since), so read the counts as growth of the public agent web, and the shares as the like-for-like comparison:
| 7 Oct | 1 Nov | Change | |
|---|---|---|---|
| Hosts checked | 18,155 | 27,559 | +52% |
| Agent Cards | 1,850 | 2,686 | +45% |
| Share of reachable hosts with a card | 10.7% | 10.1% | flat |
| Valid A2A 1.0 cards | 605 | 984 | +63% |
| Signed cards | 75 | 162 | 2.2× |
| Verified signatures | 48 | 65 | +35% |
| AI Catalogs | 696 | 1,485 | 2.1× |
| Catalogs with trust manifests | 122 | 206 | +69% |
| MCP server cards | 465 | 1,303 | 2.8× |
Publishing a card grew only as fast as the frame. Doing it properly grew faster: valid 1.0 cards, signatures, catalogs and server cards all outpaced it. Verification did not: signed cards more than doubled, verifiable ones rose by a third.
What to do with this
If you publish an agent:
- Move the card to the 1.0 shape: a
supportedInterfaceslist with a URL, binding and protocol version for each interface. - Declare your security schemes, and mark them required if they are. A client can only send what the card asks for.
- If you sign the card, publish the key: a
jkuon your own site that serves the key set. A key ID with nowhere to fetch it can't be checked. - Publish an
ai-catalog.jsonso one fetch tells a client everything you offer.
If your agents connect to other agents: treat an unsigned or unverifiable card as an unverified claim, prefer agents whose card verifies against a key on their own domain, and check the declared authentication against what the endpoint actually requires.
HVTracker keeps a signed public record of agent software; the per-host results behind this post are in the data below, and every listed project's provenance is on its profile page.
Method and limits
- Frame: 27,559 hosts from three public sources, recorded per host: the host of every remote URL in the official MCP registry at its latest versions (26,800 hosts, from 36,458 URLs across 54,149 servers), the homepage host of every project listed on HVTracker (758), and Agent Card hosts from the A2A Registry's public sitemap (44; its full listing needs an account). Shared platforms such as github.com are excluded, because a well-known path there belongs to the platform. 26,484 hosts were reachable.
- Fetches:
/.well-known/agent-card.json,/.well-known/agent.jsonand/.well-known/ai-catalog.jsonper host, and<url>/server-cardfor up to 20 MCP remote URLs per host. robots.txt was honoured (2,427 hosts block part of it); requests carried an identifying User-Agent, one worker per host with a pause between requests, a 10-second timeout and a 512 KB cap. No credentials, no API calls. - Validity checks the fields A2A 1.0 marks as required (
specification/a2a.proto), including each interface's URL, binding and protocol version. It doesn't validate types or values beyond that. - Signatures are checked best-effort. A2A 1.0 signs the card after removing default-valued fields per its protobuf schema; without that schema we try the card as served and two approximations of the stripped form, fetch the key from the
jkuit names or use an embedded key, and count a card as verified only when one form checks out. - Limits: cards are self-declared, so a declared scheme isn't proof of enforcement and a missing one isn't proof of an open endpoint. The frame leans on MCP hosts, and the A2A Registry part is a sample. These are point-in-time figures from one run.
Data: hosts.csv (one row per host that serves an Agent Card, an AI Catalog or an MCP server card: 3,978 hosts) and metrics.json (every figure above, per source, with the October comparison), under CC BY 4.0. The study runs monthly as a GitHub Action with a2a_card_study.py; a2a_census_publish.py turns a run into these files. If your host is in the data and you think a result is wrong, open an issue.
Check what you're about to connect
Paste a repository, package or MCP server and get its provenance, evidence coverage and a signed verdict.
Verify an agent or MCP server