All posts
· 9 min read ·

The AI agent directory landscape, mid-2026: 134,000 listings, 1,200 live agents

PulseMCP lists 22,282 MCP servers, Glama 59,995, chiark tracks 11,291 — and the union of the public sources holds 134,270 entries. Probe them and 1,186 answer a live protocol request. A field guide to who lists, who grades, and who probes.

Ask the public directories of the agent web how big it is and you get four different answers. PulseMCP lists 22,282 MCP servers, updated daily. Glama lists 59,995. chiark.ai tracks 11,291 agents across A2A and MCP and reports 9,776 of them online. Agenstry's own crawl — the union of eight-plus public sources, deduplicated — knows about 134,270 distinct listings.

There is a fifth number, and it is the more useful one. Of those 134,270 listings, the number that answered a live protocol request — an A2A JSON-RPC message/send or an MCP initialize + tools/list — in our most recent probe cycle is 1,186. Under 0.9%.

That gap is not a scandal, and reading it as one gets the ecosystem wrong. But it is the single most useful thing to understand about agent discovery in mid-2026: the directories disagree about the size of the agent web because counting listings and testing endpoints are different measurements, and almost everyone publishes the first.

Who lists, who grades, who probes

The directory landscape sorts into three tiers by how much verification sits behind an entry.

Listing registries record that something exists. The official MCP Registry is the canonical example: a cursor-paginated API of server manifests, submitted by publishers. Smithery and PulseMCP aggregate MCP servers at larger scale; a2aregistry.org does the same for A2A agents. A listing tells you a manifest was published. It does not tell you whether the endpoint behind it resolves, speaks its protocol, or still matches the manifest.

Grading directories add metadata-derived signals. Glama attaches A–F grades for license, quality, and maintenance to its ~60,000 entries. This is genuinely more information than a bare listing — but the inputs are repository metadata, not runtime behaviour. A server can carry an A grade and be unreachable; the grade measures the project, not the deployment.

Probing indexes test the endpoint. chiark.ai probes its tracked set every 30 minutes across three tiers — availability, conformance, performance — and publishes a 0–100 operational score with the methodology in the open. Agenstry probes on a tiered cadence, computes a conformance score across nine criteria, and layers identity on top: JWS card signatures, DNS ownership verification, GLEIF LEI linkage, and x402 payment metadata extracted during the probe. The two indexes disagree on plenty — corpus construction, scoring weights, what "online" means — but they share the premise that an entry should be evidence about a running endpoint, not a copy of its README.

The tiers are complements, not competitors. Listing registries are where publishers declare intent; probing indexes are where consumers check reality. The problem is that the ecosystem's headline numbers almost all come from the first tier.

What probing actually finds

Here is what the funnel looks like when you run it, from our snapshot of 2026-07-22:

The agent-web discovery funnel, July 2026 Four horizontal bars showing the collapse from 134,270 directory listings to 7,392 endpoints with a valid, introspectable manifest (5.5 percent), to 1,186 endpoints answering a live protocol request (0.9 percent), to 38 cryptographically signed agent cards (0.03 percent). Each bar is proportionally narrower than the one above; the last two are rendered as thin slivers for legibility. listed across public sources 134,270 valid, introspectable manifest 7,392 — 5.5% answers a live protocol request 1,186 — 0.9% cryptographically signed card 38 — 0.03%

Split by protocol, from the same snapshot: on the A2A side, 5,148 indexed agents reduce to 3,623 with a currently-valid agent card, of which 340 answer a live JSON-RPC request correctly — and 38 carry a JWS-signed card. On the MCP side, 129,122 known server names reduce to 3,769 that could be introspected as remote endpoints, of which 846 complete a live initialize handshake and return a tool list.

Three honest caveats before anyone quotes "99% of the agent web is dead."

First, most of the collapse is definitional, not failure. Of the 129,122 MCP names, 125,353 are metadata-only: package-index entries, many of them stdio-local servers that are designed to run on a developer's machine rather than as remote endpoints. A local server that installs and runs fine is not "dead" because a crawler can't reach it; it is simply not a remote service, and directories rarely distinguish the two.

Second, the bar matters. chiark reports 86.6% of its tracked set online; we report under 1% of our raw union answering live. Both are accurate. Their corpus is pre-filtered to endpoints worth tracking and their bar is availability; our denominator is every name any source has ever published and our bar is a correct protocol response. Between HTTP-reachable and protocol-conformant sits a large population — we count 1,191 A2A endpoints that respond with the wrong shape — and every directory draws its line somewhere different. This, more than crawl quality, is why five registries publish five different counts.

Third, the union is inflated by design. Registries barely overlap — we measured the overlap matrix last quarter and found the diagonal mostly empty — so a registry-of-registries inherits every source's dead weight along with its unique finds.

With the caveats granted, the residual is still striking: the population of agents that a router could actually call right now — reachable, conformant, current — is three orders of magnitude smaller than the population of things listed. And the trust layer on top of that is thinner still: 38 signed cards today, up from zero of 1,215 valid cards this spring. Signing is growing from nothing, slowly.

What to demand from a directory in 2026

If you are building on top of agent discovery — a router, an orchestrator, a marketplace, an enterprise allow-list — the practical question is not "which directory has the most entries." It is "which claims can I verify." A reasonable checklist:

  • Probe evidence, with a cadence. An uptime percentage means nothing without knowing how often it was measured and over what window. 30-day windows at a 30-minute-to-hourly cadence are the current standard among probing indexes.
  • Protocol conformance, not HTTP 200. A reachable host that returns the wrong shape will break your integration exactly as thoroughly as a dead one. The wrong-shape population (1,191 endpoints in our A2A corpus alone) is invisible to availability checks.
  • Latency distributions. A single "fast" claim hides tail behaviour; per-endpoint latency measured at probe time is the minimum for routing decisions.
  • Payment metadata. x402-priced endpoints are a growing slice, and a directory should detect them from runtime behaviour — parsed payment headers on a live 402 — not from manifests. Our paid-API listing is built from exactly that extraction.
  • Identity signals. Card signatures, DNS ownership proof, legal-entity linkage. Thin adoption today, but it is the only defence against listing-squatting as the economic stakes rise.
  • A queryable API with constraints. Both probing indexes now expose this: filter agents by minimum uptime, maximum latency, or payment support and get the evidence back per row, rather than scraping a leaderboard. (Ours: min_uptime, max_latency_ms, payment_enabled on /api/search.)

None of this makes listing registries obsolete — publication has to happen somewhere, and the official MCP Registry moving to a single canonical publication API is good for everyone downstream. It means the division of labour is becoming legible: registries for declaration, indexes for verification.

What we're watching

Whether the headline numbers converge. The honest version of the agent web's size is a funnel, not a count, and directories that publish live-rates alongside listing totals will make the single-number comparisons obsolete. Whether card signing moves — 38 of 3,623 is 1%, up from zero in May, and it is the layer everything else eventually leans on. And whether the wrong-shape population shrinks: it is the best available proxy for how much of the ecosystem tracks spec releases, with the 2026-07-28 MCP specification landing now.

The agent web is real, and it is smaller than its directories say. Both halves of that sentence are good news, if you measure them.

Sources

← Back to blog Agenstry