| Who is calling | Requests | MCP clients, by name |
|---|---|---|
| Answer engines — ChatGPT-User, PerplexityBot, OAI-SearchBot, Claude-User | 17.0M | python-httpx, undici, opencode, Codex, node |
| Training crawlers — ClaudeBot, GPTBot, Meta-ExternalAgent | 10.5M | claude-code, go-http-client, Claude-User, Cursor |
| Resellers, search APIs and scrapers | 11.4M | MCPWatch, MCPScoringEngine, mcp-rugpull-research, agentprobe |
Discovery Is the First Governance Control
The MCP landscape, measured across 31,229 API providers
Tyk LEAP ON TOUR · New York · 21 October 2026
Kin Lane — API Evangelist · apis.io
Future of MCP Governance
“Every call is an ungoverned path unless something inspects it.”
That is Tyk’s line for today, and it is right. But nothing can inspect a path it has not found. Discovery comes first.
Fifteen minutes of what the public market actually looks like — measured, not announced — so the panel and your tables have something real to argue with.
Who is talking — thirty seconds, not a resume
Sixteen Years, One Instrument
API Evangelist since 2010. Presidential Innovation Fellow. Chief Evangelist at Postman. API governance across Bloomberg’s estate.
Today: APIs.io — 31,229 providers and 154,842 APIs profiled and scored on two instruments that never touch each other: the Kin Score (governance) and Agent Readiness (nineteen dimensions).
I also run two public MCP servers of my own, published to the official registry with OAuth 2.1 in front of them — so everything I say about the market, I have had to do.
For the business side of the room
What MCP Is, in One Slide
- Model Context Protocol
- A standard way for an AI agent to ask a system “what can you do for me?” — and then do it.
- A server publishes
- Tools (actions), resources (things to read) and prompts. A client — Claude, ChatGPT, Cursor, Copilot, your own agent — connects, lists the tools, calls them.
- Underneath
- An HTTP endpoint calling your APIs, with OAuth 2.1 for authorization and runtime discovery of the authorization server. It inherits everything your APIs already are.
Measured on my own properties · 13 Aug – 19 Oct 2026 · 68 days
Agents Are Already Here
35.8%
of 113M requests came from declared AI agents and crawlers
2.7×
developer-agent calls in two weeks — Claude Code, Codex, Cursor, opencode working against the API
1.3M
direct MCP calls to my two servers · 414 distinct clients in one day
APIs.io catalog · build of 2026-10-17
“MCP Server” Is Three Different Numbers
6,968entries — anything a provider says about MCP, machine-readably
4,041real servers — a manifest with a server behind it
3,321hosted endpoints an agent can actually open
| How it is delivered | Entries | Share |
|---|---|---|
| Hosted endpoint — a URL an agent can open | 3,321 | 48% |
| Candidate descriptor — declared, nothing reachable behind it | 2,774 | 40% |
| Documentation link — a page about a server | 691 | 10% |
| GitHub repo — run it yourself | 178 | 3% |
| Providers with any MCP entry at all | 6,745 of 31,229 | 22% |
The reachable third, probed
Reachability Is Not Governance
836of 2,548 probed remote endpoints answered an anonymous initialize — 33%
1,463refused with an auth challenge — 57%
160dead hosts, or not an MCP endpoint at all
A random eighty, by hand: 26 completed the handshake with no credential,
32 refused with an HTTP auth error, 15 refused inside JSON-RPC, 5 were not MCP — and
2 were uninstantiated templates, shipped with the placeholder still in the URL:
https://<JFROG_PLATFORM_URL>/mcp · /mcp/{ACCOUNT_ID}
What the MCP spec's own authorization model depends on · 31,004 scored providers
The Authorization Gap
6.7%publish RFC 9728 protected-resource metadata
5.2%let a client register itself (RFC 7591)
24.9%hand out a bearer token or API key and nothing else
| Dimension | Providers | Share |
|---|---|---|
| Protected resource metadata (RFC 9728) — served and verified | 2,075 | 6.7% |
… served specifically at the /mcp path | 140 | 0.5% |
| Dynamic client registration (RFC 7591) — no portal form | 1,617 | 5.2% |
| Delegated user identity — a token scoped to the human, served | 2,530 | 8.2% |
| Auth clarity: a served OIDC / OAuth discovery document | 2,760 | 8.9% |
| Auth clarity: bearer token or API key only | 7,710 | 24.9% |
| Sender-bound credentials (mTLS, DPoP, HTTP signatures) | 21 | 0.07% |
4,408 servers with a captured tool list
Curated, or Generated?
12median tools per server
182servers with more than 100 tools — the largest, 1,029
42%of servers with five or more tools are one-tool-per-operation generators (1,535 of 3,620)
- Platform-served
- 643 of the 3,677 providers credited for an MCP server have it served by a docs platform — ReadMe, Mintlify, Gram — rather than from their own host.
- The pattern
- Auto-generating an MCP server from an OpenAPI is the new auto-generating an SDK: a tool per operation, and now the model chooses between 300 of them with its context window.
What providers actually serve · 10,809 providers probed at their own domains
Discovery Is Federated, and Nobody Sees All of It
| Discovery document | Providers serving it |
|---|---|
/.well-known/oauth-authorization-server (RFC 8414) | 2,248 |
/.well-known/oauth-protected-resource (RFC 9728) | 1,578 |
/.well-known/api-catalog (RFC 9727) | 702 |
/.well-known/agent-card.json (A2A) — 490 graded, 377 conformant | 629 |
/llms.txt | 573 |
/.well-known/mcp/server-card.json · /.well-known/mcp.json | 186 · 118 |
| Official MCP Registry (registry.modelcontextprotocol.io, 19 Oct) | 46,293 · 30,018 remote |
Two scores that never touch · 31,005 providers · 2026-10-17
Readiness Is a Layer. It Does Not Rise Above Its Foundation.
r = 0.870governance composite × agent readiness
94.9%score readiness at or below their governance
1.4%reach Agent-Native · 62.0% have no agent signal at all
| Governance band | Providers | Reaching Agent-Native |
|---|---|---|
| Exemplar | 434 | 21.9% |
| Strong | 1,251 | 11.5% |
| Developing | 3,387 | 4.5% |
| Thin | 4,581 | 0.7% |
| Emerging | 7,469 | 0.1% |
| Minimal | 13,883 | 0.0% |
The negative finding
An MCP Server on an Ungoverned API Does Not Move the Number
3.9%declare an idempotency key on writes
7.3%document what can be undone
0.6%offer a dry-run on destructive operations
| What actually moves readiness | Share of 31,004 | Verified at the contract |
|---|---|---|
| Rate-limit signaling | 32.1% | 489 |
| Stable error semantics an agent can branch on | 20.3% | 2,725 |
| Request / response examples | 15.5% | 2,088 |
| Documented reversibility — what can be undone | 7.3% | 234 |
| Idempotency keys on writes | 3.9% | 502 |
| Dry-run / simulate on destructive operations | 0.6% | 167 |
| MCP server | 11.9% | 1,785 |
New York's industries · APIs.io membership by tags · 2026-10-17
The Server Is the Easy Part
4.3%of banks are agent-native — ahead of the AI companies
2.4×the MCP shipping rate at AI companies, with no more readiness to show for it
| Industry | Providers | Agent-Native | Any MCP server | Hosted MCP |
|---|---|---|---|---|
| Banking & Open Banking | 910 | 4.3% | 10.9% | 9.3% |
| Payments & Money Movement | 1,779 | 5.8% | 15.7% | 13.0% |
| Financial Services | 1,252 | 2.9% | 9.4% | 7.7% |
| Insurance | 700 | 0.9% | 7.1% | 6.6% |
| Healthcare | 2,313 | 0.4% | 5.4% | 4.9% |
| Developer Tools & DevOps | 2,994 | 2.1% | 25.6% | 18.2% |
| Artificial Intelligence | 6,303 | 2.4% | 26.1% | 22.0% |
How banking met this moment first · graded at the contract · 2026-10-17
Standards Stabilize an Industry
| Cohort | Providers | Agent-Native | Ready or Native | Median readiness |
|---|---|---|---|---|
| Conform to PSD2 / Open Banking | 178 | 9.6% | 47.2% | 27.4 |
| Banking at large | 910 | 4.3% | 23.1% | 3.2 |
| Conform to FHIR | 102 | 3.9% | 38.2% | 21.8 |
| Healthcare at large | 2,313 | 0.4% | 6.0% | 0.0 |
| Conform to FAPI (financial-grade OpenID) | 97 | 3.1% | 53.6% | 31.4 |
| Conform to CAMARA (telecom) | 12 | 16.7% | 41.7% | 27.7 |
| Telecom at large | 1,395 | 3.1% | 23.2% | 16.2 |
Claimed is not conformant. FHIR: 1,093 declare, 102 conform · PSD2 / Open Banking: 570 declare, 178 conform · FAPI: 741 declare, 97 conform.
Monday
A Governance Program That Starts With Discovery
- Inventory the three numbers for your own estate — declared, exists, reachable. The delta is your shadow MCP.
- Publish discovery on your own domain —
/.well-known/api-catalog, protected-resource metadata, a server card. Cheap, rare, and what every registry and scanner reads first. - Own the server's design — few tools, use-case shaped, annotated read vs write. A platform-generated server is a start, not an end.
- Put identity in the path — token exchange so the agent carries the user's identity, not the integration's key. Log the delegated subject.
- Fix the API underneath where readiness is bounded — idempotency, error schema, rate-limit headers, reversibility.
- Measure it on the same instrument every quarter. Nothing moves the number except what you publish — which is exactly what makes it worth moving.