snapshots

## What this dataset is

One record for every remote MCP server endpoint listed in the official MCP
Registry (registry.modelcontextprotocol.io) at capture time — 112 endpoints —
each probed once from a single vantage point, with the result recorded exactly
as measured.

For each endpoint the capture records whether the server answered an
anonymous MCP `initialize` handshake, how it was classified, and — when the
server completed the handshake — a fingerprint of its tool contract: the tool
names, a SHA-256 hash of each tool's canonical input schema, and one
whole-contract hash over all of them. That makes a contract change a single
hash comparison, and shows exactly which tool's schema moved.

Capture set in this version: **2026-10-21, 07:07–07:19 UTC**, measured from
New York, US (DigitalOcean nyc1).

Result mix for this set: 37 LIVE · 30 AUTH_GATED · 44 NO_RESPONSE · 1 DEAD.
Tool-contract snapshots are included for the 32 endpoints that completed the
handshake and served a `tools/list`.

## How an endpoint is classified

Each endpoint receives one anonymous MCP `initialize` request (JSON or SSE
transport), followed by `tools/list` when the handshake succeeds:

- **LIVE** — the server completed the handshake. If `tools/list` could not be
completed, the record carries a note saying the contract is unknown and no
snapshot is included.
- **AUTH_GATED** — the server demands authentication (HTTP 401/403 with a
Bearer challenge or an RFC 9728 protected-resource response).
- **NO_RESPONSE** — the endpoint answered but not with a usable MCP result
(wrong status, malformed body, redirect, timeout, or an auth response with
no standard challenge). The note field says which.
- **DEAD** — the endpoint could not be reached at all (DNS failure or an
HTTP 404/5xx on the handshake).

## Schema

One JSON object per line (JSONL). One record per endpoint per capture set.

| Field | Type | Present | Meaning |
|---|---|---|---|
| `endpoint_url` | string | always | The endpoint URL as listed in the registry. |
| `captured_at` | string (UTC, ISO 8601) | always | When this endpoint was probed. |
| `vantage` | string | always | Where the probe was measured from. |
| `status` | string | always | `LIVE`, `AUTH_GATED`, `NO_RESPONSE`, or `DEAD` (see above). |
| `elapsed_ms` | integer | always | Handshake round-trip time in milliseconds. |
| `tool_count` | integer | snapshot rows | Number of tools in the captured contract. |
| `tools_sha256` | string | snapshot rows | SHA-256 over the canonical per-tool schema hashes — the whole-contract fingerprint. |
| `tools` | array of strings | snapshot rows | Tool names, sorted. |
| `per_tool_sha256` | object | snapshot rows | Map of tool name → SHA-256 of that tool's canonical input schema. |
| `note` | string | when recorded | The probe's own note for this record (e.g. why a response was not usable, or that a live server's contract could not be read). |

Tool input schemas are represented by their SHA-256 fingerprints
(`per_tool_sha256`), not reproduced in full.

## Versions and diffs

Each published version is one complete capture set. Diffing two versions
shows, endpoint by endpoint: status changes (a server that went dark, moved,
or started demanding auth), tool additions and removals, and — via
`per_tool_sha256` — exactly which tools' schemas changed, with
`tools_sha256` flagging any contract change in one comparison.

## License and price

License: **CC-BY-4.0**. This dataset reads free.