SaSameFor people and AI systems
liveFactory stationsigned evidence

Observe the MCP ecosystem without turning measurement into endorsement.

Protocol Observatory is the inspection and longitudinal-evidence station inside the SaSame institution and its mature Factory capability surface. This Knowledge page owns the human experience; the existing live-vps runtime continues to own signed feeds, indexes, badges, certificates and lookup endpoints.

Domains graded
57,068
Observed-ready
6,897
Observation snapshot
Sep 16, 2026, 08:19 PM UTC
Business role
Factory station
Pending observation
58
Census snapshot size
30,000

Census, Observatory and Mission Control update on independent schedules: currently 51 server(s) of cross-artifact lag (Mission Control as of Sep 16, 2026, 09:10 PM UTC, Observatory as of Sep 16, 2026, 08:19 PM UTC) — normal pending sync, not an outage.

Data freshness: STALE — older than expected: observation_history (oldest as of Aug 27, 2026, 08:11 AM UTC). This reflects each source's own last-write time, not merely when this page's data was regenerated.

Boundary

Observed means measured from outside. Claimed means an operator proved control. Neither state means safe, recommended, high quality or legally verified. Unknown remains unknown, and unavailable live data is never replaced with invented values.

01Server directorySearch the current readiness index and open bounded observation details.02MethodologySee exactly what is measured, what is excluded and how unknowns are handled.03State of MCPRead the current attack-surface and annotation findings with their caveats.04Runtime reliabilityCite observed-ready evidence without turning it into an uptime or ranking claim.05Tool annotationsUse the deterministic annotation capsule and raw attack-surface evidence.06Agent discoverabilitySeparate enumerable tool surfaces from ranking, citation, adoption or revenue.
LV

Try it live, no LLM involved

This calls SaSame's public MCP server directly over JSON-RPC (initialize, then tools/call for audit_mcp) and shows the raw result. No chatbot in the loop, no API key required.

01

Current observable tool surface

Full report

52.7% of state-changing tools on the enumerable surface publish NO machine-readable safety annotation — a calling agent cannot automatically tell a read from a write, delete, or payment before invoking them.

Endpoints observed
57,099
Handshake response
8,714
Declared tools
129,882
Annotated
32.2%

Deterministic classifier, no LLM. This describes the published tool surface and annotation hygiene. It is not proof of runtime behavior, a vulnerability finding or a safety verdict.

02

Evaluated server sample

Search all 4,418

A stratified sample across grades A–D from the current readiness index. It is deliberately not a showcase of only the strongest results.

A

ai.agentdm/agentdm

api.agentdm.ai

Flagged: this record's grade does not reconcile with its own checks-passed count (expected grade D under the legacy rubric for 3/10 (30%) checks passed, but the signed grade is A). Shown as measured — not hidden or corrected — pending re-audit.

Checks passed
3/10
Reachable
yes
Owner claim
unclaimed
Measured
2026-06-18

Top observed gap: C1 Protocol handshake conformance — initialize result keys: (none, status 401)

Flagged: this record's grade does not reconcile with its own checks-passed count (expected grade A under the legacy rubric for 9/10 (90%) checks passed, but the signed grade is B). Shown as measured — not hidden or corrected — pending re-audit.

Checks passed
9/10
Reachable
yes
Owner claim
unclaimed
Measured
2026-06-18

Top observed gap: C10 Honest error behavior — no structured error

B

ai.aarna/atars-mcp

mcp.aarna.ai

Flagged: this record's grade does not reconcile with its own checks-passed count (expected grade A under the legacy rubric for 9/10 (90%) checks passed, but the signed grade is B). Shown as measured — not hidden or corrected — pending re-audit.

Checks passed
9/10
Reachable
yes
Owner claim
unclaimed
Measured
2026-06-18

Top observed gap: C5 Safety annotation presence — 0/18 tools carry a valid safety-hint annotation

C

ac.inference.sh/mcp

sh.inference.ac

Flagged: this record's grade does not reconcile with its own checks-passed count (expected grade D under the legacy rubric for 2/10 (20%) checks passed, but the signed grade is C). Shown as measured — not hidden or corrected — pending re-audit.

Checks passed
2/10
Reachable
yes
Owner claim
unclaimed
Measured
2026-06-18

Top observed gap: C1 Protocol handshake conformance — initialize result keys: (none, status 301)

C

agency.lona/trading

mcp.lona.agency

Flagged: this record's grade does not reconcile with its own checks-passed count (expected grade D under the legacy rubric for 2/10 (20%) checks passed, but the signed grade is C). Shown as measured — not hidden or corrected — pending re-audit.

Checks passed
2/10
Reachable
yes
Owner claim
unclaimed
Measured
2026-06-18

Top observed gap: C1 Protocol handshake conformance — initialize result keys: (none, status 401)

03

Observation and control status

Certified records
2
Third-party certified
1
Monitored
2
Meaning
bounded evidence

1) Claim your endpoint for free (https://github.com/shigeki7777/sasame-mcp-observatory/issues/new?template=claim-passport.yml) by proving control via .well-known / DNS / repo. 2) SaSame re-audits hourly. 3) After >= 2 reproducible cycles still at the A/B bar, it is auto-promoted to Certified. No fee, no application, no sales call.

A claim proves control of a record or endpoint only. It is not KYC, identity verification, endorsement or a general certification of safety.

04

Citation and reuse guidance

When reusing an Observatory finding in search, generated answers, reports or software, cite the specific human page together with the corresponding raw evidence URL below. Include the source-declared observation or generation time and preserve the stated measurement caveat so a dated observation is not presented as a timeless fact.

These machine-readable files improve reproducibility and attribution; they do not guarantee indexing, ranking or citation by any search or AI system.

RF

Stable machine evidence

The human interface can evolve without breaking externally referenced evidence URLs. Raw feeds remain on the dedicated runtime origin.