SaSame MCP Observatory · Directory Pre-Flight

Check your MCP, fix the mechanical blockers, then claim its Readiness Passport.

SaSame's free pre-flight measures the mechanical readiness that agents and directories care about: handshake, tools/list, schemas, descriptions, annotations, real-content behavior, and honest errors. After the grade, claim the owner-controlled Passport so agents can identify the operator-controlled record.

Check → grade → claim Observatory home The open standard

What a pre-flight catches — and what it can't

✅ Caught (mechanical, measurable)

  • Tool annotations — every tool needs a title + readOnlyHint or destructiveHint (missing annotations are ~30% of Claude rejects).
  • Typed input schemas (untyped / invalid schema).
  • Description clarity & distinctness (vague or templated descriptions).
  • Liveness / reachability over HTTPS; handshake conformance.
  • Graceful errors (no generic 500/400 on malformed input).
  • Anti-ghost — a read-only tool actually returns real content.
  • Promotional / overly-generic tool names (best, official, helper…).

❌ Out of scope (you handle these)

  • Privacy policy — required in your public docs; missing/incomplete is an immediate reject on Claude. A probe can't verify its content.
  • Identity / business verification (publishing under an unverified name is rejected on ChatGPT).
  • OAuth setup & callback allow-listing.
  • Data minimization / undisclosed PII in tool responses.
  • Prohibited categories (financial/crypto transfer, standalone media generation, ad-primary).
  • Read/write separation into distinct tools (design).

Authoritative requirements live with the directories themselves: Claude Connectors submission · Anthropic Software Directory Policy · ChatGPT Apps submission.

How to run it (free, handshake-only)

Point either tool at your own public MCP endpoint. The probe is a legitimate MCP handshake (initialize + tools/list + one read-only tool call) — no auth-bypass, no payment, no data written to your server.

Public MCP endpoint

https://live-vps.sasame.online/public-mcp
ToolReturns
audit_mcp(url)A→D grade, per-criterion pass/evidence, the single biggest gap, plus a machine-readable preflight block (PF1–PF3).
readiness_report(url)The full Markdown report: every criterion with a fix, a Directory Pre-Flight section, and an ed25519-signed MCP-Ready certificate at grade A/B.
verify_mcp_ready(url)A portable, offline-verifiable signed certificate of the same audit.

Prefer the command line?

Run the same readiness audit locally in one command — no MCP client, no install, no key (open-source, MIT):

npx mcp-readiness https://your-server.example/mcp

Same A→D grade and directory pre-flight as the hosted tools. The latest CLI also prints the claim path after the grade. On npm: mcp-readiness.

Pre-flight checks (advisory layer)

IDCheckMaps to
PF1Every tool has a title + readOnly/destructive hintClaude & ChatGPT: missing/incorrect action labels — a top reject cause
PF2Specific, non-promotional tool namesChatGPT: generic/comparative names may be rejected
PF3Privacy-policy URL surfaced over MCP (soft reminder)Claude: missing policy = immediate reject (verify in your public docs)

These map to the canonical, reproducible 10-criterion standard (C1–C10) but are reported as a separate advisory layer — a third party's directory policy never alters the reproducible grade.

After the grade: claim the Passport

If the endpoint is yours, call claim_start(url), publish the returned token through .well-known or DNS, then call claim_confirm. The result is an owner-controlled, agent-readable Readiness Passport. It is still measured-facts only: no safety verdict, no paid ranking, no KYC, no custody.

Open GitHub claim template Agent instructions

Why trust the result

Every criterion is bound to the MCP spec, the registry schema, cryptography, or direct measurement — not taste — so anyone can re-run the audit and reach the same booleans. SaSame reports verification status only; it never issues a safety, quality, or "worth-using" verdict on any named server. SaSame's own public MCP server self-audits at grade A and passes PF1/PF2 (and PF3 is a standing reminder it applies to itself).