Ruleset 2026-07-28, re-diffed against the final 2026-07-28 specification text on 2026-08-03: every verdict-bearing rule unchanged.
31 of 11,929 reachable MCP servers are fully ready for the 2026-07-28 spec
On 2026-09-13 we probed every one of the 31,538 servers in the official registry against the Model Context Protocol's 2026-07-28 revision. Of the 11,929 reachable through a conformant keyless initialize, 31 pass all four required-conformance signals, against 29 in the 2026-09-09 census. That is 0.3% of the reachable population, so the honest reading is that migration has started, not that it has progressed. The counterweight is real either way: 63.3% already tolerate handshake-free stateless calls, the heart of the change. Dataset: CC BY 4.0, tag v2026.09.13-7c7cb7f.
What moved since 2026-09-09
The leading indicators moved further than the ready count, and in the same direction, while the reachable population grew 15.4%. That is what makes the 31 look like the front of a wave rather than noise.
| Signal | 2026-09-09 | 2026-09-13 |
|---|---|---|
| fully ready | 29 | 31 |
| server/discover implemented | 393 | 420 |
| routing headers enforced | 280 | 307 |
| three of four signals | 314 | 327 |
| at-risk | 1,881 | 1,898 |
At-risk is flat in absolute terms while its share of a larger population falls. Nothing is being fixed on that surface; new arrivals are simply landing elsewhere.
Readiness verdicts
| Verdict | Servers | Share of reachable | Meaning |
|---|---|---|---|
| ready | 31 | 0.3% | passes all four required-conformance signals |
| needs-migration | 10,000 | 83.8% | reachable, answers the old handshake flow, missing the new requirements |
| at-risk | 1,898 | 15.9% | on a removed or deprecated surface (required sessions, HTTP+SSE transport, experimental Tasks) |
The remaining 19,609 registered servers carry no runtime verdict: package-only, auth-gated, or down, with no endpoint a keyless probe can measure. That is why every percentage on this page uses the 11,929 denominator.
The four required signals, individually
Accept handshake-free calls
7,555 (63.3%)
The stateless core is the widest-adopted piece by far
Session-independent
9,597 (80.5%)
2,332 still mint Mcp-Session-Id; 1,709 hard-require it
Implement server/discover
420 (3.5%)
The new mandatory discovery RPC, still nearly absent
Enforce routing headers
307 (2.6%)
Reject a mismatched Mcp-Method header as the revision requires
327 servers pass three of the four signals: one condition away from ready. The bottleneck is not the architectural change; it is the two new mechanical requirements, server/discover and header enforcement, which ship by default in the new SDK betas.
What versions servers speak today
| Negotiated protocol version | Servers | Share of reachable |
|---|---|---|
| 2025-06-18 | 8,586 | 72.0% |
| 2025-03-26 | 2,467 | 20.7% |
| 2024-11-05 | 618 | 5.2% |
| 2025-11-25 | 124 | 1.0% |
| 2026-07-28 | 123 | 1.0% |
| 2025-11-05 | 3 | 0.0% |
| 2026-09-02-agent-first-v1 | 1 | 0.0% |
123 servers already answer with 2026-07-28. 618 still negotiate 2024-11-05, a version three revisions back. Servers that completed initialize without exposing a version are not in this table.
The registry grew 117% while nobody migrated
Between our 2026-07-02 baseline census and this one, the official registry grew from 14,559 to 31,538 servers, 117% in twenty days, while health rates stayed flat (80.6% healthy). The ecosystem is expanding on the old protocol faster than it is migrating off it. Both dated editions are public in the dataset repository, so this comparison is reproducible, not asserted.
If you run an MCP server
- Check your own server: the free check form resolves your registry entry, or run akashi locally:
akashi check your-namespace/your-serverprints a per-signal spec 2026-07-28 line. - If you are on an SDK, upgrade to the 2026-07-28 beta (Python, TypeScript, Go, and C# betas are out): server/discover, header enforcement, and the stateless core come with it.
- If you issue Mcp-Session-Id, plan the move to server-minted handles passed as tool arguments; required sessions are the single largest at-risk surface in this census.
FAQ
- Is any MCP server ready for the 2026-07-28 spec?
- 31 of the 11,929 reachable servers (0.3%) pass all four required-conformance signals (server/discover, handshake-free stateless calls, session independence, routing-header enforcement). 327 servers are exactly one condition away, and 63.3% already accept handshake-free calls, the heart of the new revision.
- Will MCP servers break on July 28, 2026?
- Not on day one. The revision ships with a formal deprecation policy and a minimum twelve-month window for legacy versions, and clients keep speaking older protocol versions. Readiness here measures conformance with the new revision, not whether a server stops working the day it publishes.
- How many MCP servers still depend on sessions?
- 2,332 of the 11,929 reachable servers (19.5%) still mint the removed Mcp-Session-Id header, and 1,709 hard-require it: they reject requests without a session and cannot serve new-spec clients without change.
- Why is the denominator 11,929 and not all 31,538 registered servers?
- Because runtime readiness can only be measured against a server you can actually reach. Most registered servers are package-only (npm, PyPI, Docker) with no hosted endpoint to probe keylessly. Quoting a readiness percentage over the registered total would silently count unmeasurable servers as unready.