date string | mcp_total int64 | mcp_open int64 | mcp_auth int64 | mcp_behind_latest int64 | mcp_v_2024_11_05 int64 | mcp_v_2025_03_26 int64 | mcp_v_2025_06_18 int64 | mcp_v_2025_11_25 int64 | mcp_tokens_p50 int64 | mcp_tokens_p90 int64 | mcp_tokens_max int64 | mcp_prm_missing int64 | mcp_prm_unreachable int64 | a2a_total int64 | a2a_cards int64 | a2a_schema_0_3 int64 | a2a_schema_1_0 int64 | a2a_registry_mismatch int64 | a2a_signed int64 | a2a_invalid_basic int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-10-20 | 338 | 117 | 173 | 36 | 3 | 7 | 26 | 81 | 1,644 | 10,088 | 98,668 | 14 | 4 | 571 | 544 | 333 | 211 | 50 | 44 | 43 |
Agent Surface Index — aggregates
Weekly aggregate measurements of publicly reachable AI agent endpoints. One row per snapshot.
This dataset contains aggregates only. No per-endpoint records are published.
- Measurement harness and method: https://github.com/deeparchi-ai/agent-surface-index
- Full report for the first snapshot: see
REPORT.mdin that repository
What is measured
MCP endpoints (Streamable HTTP) — reachability via initialize, negotiated protocolVersion, declared capabilities, and the cost of tools/list (serialized payload length ÷ 3.6 as a token estimate, used only for within-snapshot comparison). Latency is the median of three tools/list calls. Authorization discovery follows 401 → WWW-Authenticate → Protected Resource Metadata → authorization-server metadata, with well-known URLs built per RFC 8414 path insertion.
A2A agents — .well-known/agent-card.json reachability, card schema generation (0.3 has top-level protocolVersion; 1.0 uses supportedInterfaces), required fields, presence of signatures, and whether registry metadata agrees with the card.
No tools/call is ever issued. The harness does not invoke business actions.
Columns
| Column | Meaning |
|---|---|
date |
Snapshot date |
mcp_total |
MCP endpoints attempted |
mcp_open |
Reached initialize (no auth) |
mcp_auth |
Returned 401 |
mcp_behind_latest |
Open endpoints not on the newest dated protocol version observed in that snapshot |
mcp_v_* |
Open endpoints per negotiated protocol version |
mcp_tokens_p50/p90/max |
Estimated tools/list token cost |
mcp_prm_missing |
401 endpoints without resource_metadata |
mcp_prm_unreachable |
Metadata URL present but unreadable |
a2a_total |
Registered agents attempted |
a2a_cards |
Cards retrieved |
a2a_schema_0_3 / a2a_schema_1_0 |
Card schema generation |
a2a_registry_mismatch |
Registry protocolVersion differs from card |
a2a_signed |
Cards carrying signatures |
a2a_invalid_basic |
Cards missing required fields for their schema |
Policy
- Facts, not grades. No scores, no rankings, no badges.
- No named endpoints in this dataset.
- Security-relevant counts are aggregated deliberately; affected parties are notified privately before anything is named.
- Vendors may request the record for their own endpoint.
Caveats
- A snapshot is a point-in-time observation, not an SLA. Latency depends on the network vantage point.
- Token counts are character-based estimates.
- Registry populations skew toward long-tail projects and are not a census of production deployments.
- Endpoints behind authentication are measured for discovery only; their real capabilities are unknown.
Licence
Data: CC BY 4.0. Harness code: Apache-2.0.
- Downloads last month
- 12