KanseiLink collects aggregated, anonymized usage data through the report_outcome tool. This data helps agents make informed decisions about MCP service quality.

• Service ID and success/failure status
• Response latency (optional)
• Error category (optional, predefined types only)
• Context text (optional, PII auto-masked before storage)

• Agent identity (hashed to anonymous ID)
• User personal information (auto-masked)
• Authentication credentials
• Request/response payloads from MCP services
• IP addresses of calling agents

All text submitted via context field is processed through PII masking before storage:

Policy: Raw text with PII is never persisted to disk. Masking occurs in-memory before any database write.

• agent_id (e.g. on submit_feedback) is normalized to an agent family (claude / gpt / gemini / …) or anonymous — an arbitrary identifier (email, username) can never be stored through it.
• Auto-captured error responses (kansei-link-report-hook) are PII-masked before classification; only the resulting error category is ever transmitted, never the raw response.

• The npm package (@kansei-link/mcp-server, stdio server) makes no outbound network calls by default — it serves a local bundled SQLite dataset. The single exception is opt-in: if you set KANSEI_API_KEY, the server validates that key against GET /api/validate-key (tier check only, cached 10 min; no usage data is sent).
• The hosted HTTP facade (Railway) exposes the dashboard read APIs, Stripe billing, auth/entitlements, and the opt-in report-outcome / telemetry sinks.

• Stripe webhooks (/webhooks/stripe) are verified with stripe.webhooks.constructEvent (HMAC signature) over the raw request body.
• /api/checkout accepts only price IDs configured via STRIPE_PRICE_* (a client cannot substitute an arbitrary or different valid price).
• /api/access is a low-stakes tier/expiry read keyed by email (no payment data, no content). Actual premium content is never unlockable by email alone — see /api/premium below.
• Magic-link email login (/api/auth/request-link → emailed one-time code → /api/auth/verify) issues the per-email access token by proving inbox control. Codes are stored hashed, expire in 15 minutes, are single-use, and the request endpoint answers identically for customers and non-customers (no enumeration). Outbound mail goes through Resend only when RESEND_API_KEY is configured.
• /api/portal (manage/cancel billing — high-impact) is token-gated: it requires the per-email HMAC(email, secret) token, so it cannot be used to take over or enumerate billing. The same token gates /api/keys (API key create/list/revoke).
• /api/premium serves gated article sections to (email + access token) or a valid API key of sufficient tier. The premium HTML lives only in the server DB — the public repo and the static GitHub Pages HTML do not contain it.
• API keys (kl_…) entitle MCP/HTTP clients to paid tiers. Only SHA-256 hashes are stored; the plaintext is shown once at issue time. Tier is resolved live from the subscription, so cancellation downgrades keys automatically. GET /api/validate-key returns tier only — never the owning email.
• /admin/* endpoints (including premium-content upload) require a CRAWLER_SECRET bearer token.
• CORS is scoped to a single configured origin; credentials are not exposed.

Server secrets are environment variables only (never committed): STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, STRIPE_PRICE_*, CRAWLER_SECRET, ACCESS_TOKEN_SECRET (optional — falls back to STRIPE_WEBHOOK_SECRET), RESEND_API_KEY (magic-link email delivery), EMAIL_FROM, ANTHROPIC_API_KEY, GITHUB_TOKEN. See .env.example.

• Initial score: 0.5 (neutral)
• Adjusted based on: namespace verification, community outcomes, manual review
• Scores below 0.3 trigger a warning in search results

Calculated from:

• Unique agent count (40% weight): More independent agents = more trustworthy
• Total call volume (30% weight): More data points = more reliable
• Data recency (30% weight): Fresher data = more relevant

1. Structural validation (MVP): Input schema validation via Zod
2. Statistical anomaly detection (MVP): Basic outlier detection on latency/success patterns
3. Cross-validation (planned): 3+ independent agent confirmations boost confidence
4. Human review (planned): Flagged anomalies reviewed by maintainers

Contact: security@synapsearrows.com (or open a GitHub security advisory)

KanseiLink uses io.github.kansei-link/* namespace, verified via GitHub OIDC. We do not claim domain-based namespaces until DNS ownership is confirmed.