AgentProof · Trust Centre
Trust Centre
Honest positioning. What AgentProof is and is not. What it reads, what it writes, what it claims. No exaggeration.
What AgentProof is
AgentProof is a provider-agnostic agentic-AI design-readiness lens. It scores how well an agent is set up to behave responsibly.
What AgentProof is not
AgentProof is not a runtime monitor. It does not watch live agent traffic.
What AgentProof is
Every reading is evidence-backed. Old reports are immutable so progress is real, not retroactive.
What AgentProof is not
AgentProof is not an accrediting body. It does not issue regulatory or legal approvals.
What AgentProof is
Connectors read only the read-only metadata needed for the AgentProof footprint. Microsoft Power Platform / Copilot Studio is the first real connector path.
What AgentProof is not
AgentProof does not extract business data by default. It does not require source-code access. It does not perform broad tenant scanning.
What AgentProof is
AI Radar surfaces approved intelligence updates from named sources. Every change passes a human review gate before it can affect reports.
What AgentProof is not
AgentProof does not crawl the open web. It does not auto-publish methodology updates. It does not watch live agent traffic.
What AgentProof is
Customer data stays inside the customer workspace. Row-level security enforces auth.uid() = user_id on every customer table.
What AgentProof is not
AgentProof does not use the service-role key in the customer path. It does not echo env-var values. It does not log tokens.
What AgentProof is
AgentProof is in private pilot. Workspaces are invitation-led.
What AgentProof is not
AgentProof has no checkout flow inside the product. No invoices, contracts, or quotes are generated here.
Trust Centre · claim safety
12 founder-authored sections · 0 prohibited-claim violations
Each section is plain language. The claim-safety validator scans every section for prohibited phrases (certified, ISO 27001, SOC 2, GDPR-compliant, Microsoft endorses, guaranteed safe, vulnerability scan, etc.) — when this panel shows zero violations, no Trust Centre copy makes a claim it cannot back up.
What AgentProof reads
Read-only metadata: tenant identifiers, environment list, agent list, configuration objects needed to build the canonical agent footprint. Read-only consent is granted by your tenant admin and can be revoked at any time.
What AgentProof does not read
No business records. No conversation content. No customer PII. No transactional data. No model weights. No payment data. No contract data.
Read-only connector posture
Every connector AgentProof ships requests least-privilege read-only scopes. AgentProof never asks for write or modify scopes. Tokens never leave the server.
Workspace isolation
Every workspace is scoped to its owner via Supabase RLS auth.uid() = user_id on every policy. Cross-tenant viewers never see another customer's connector state, discovery facts, reports, or improvement actions.
Demo and sample separation
Sample agents always carry a Sample or Demo badge. Sample data cannot enter a real workspace without that badge. Real workspaces start empty and only ever show real data the workspace owner has authorised.
Authentication model
Magic-link sign-in only. No password, no SSO, no third-party trackers. Local test access is available in non-production builds only and requires a server-side env flag plus a configured test user.
Data handling principles
Workspace data is stored in your hosted Supabase project. AgentProof does not send your data to any live AI provider unless you explicitly configure one. The Supabase project is expected to be EU-region when EU-region is required.
Microsoft endorsement
AgentProof is not endorsed by Microsoft. References to Microsoft Power Platform, Copilot Studio, Azure AI Foundry, or Entra ID describe public Microsoft surfaces only — AgentProof does not speak on behalf of Microsoft.
Legal and certification claims
AgentProof does not claim ISO 27001, SOC 2, GDPR, HIPAA, FedRAMP, or any other certification. AgentProof is not legal advice. AgentProof is a readiness lens — it shows what is in place and what is missing.
Revocation awareness
Any consent you grant AgentProof can be revoked at any time through your Microsoft tenant admin console (or equivalent for other providers). AgentProof refuses to operate on a revoked connector.
Audit-friendly design
Every discovered fact records a discovery_source tag (real_connector_call, simulated, user_confirmed, manual_fallback). Reports carry a version stamp and methodology version. Connector state changes are journaled.
AI provider usage boundaries
No live AI provider is contacted at any point in the public build. When live providers are configured by the operator, AgentProof uses them only to generate text narrative — the deterministic readiness score is computed without any LLM and cannot be overridden by an LLM.
Claim-safety validator: CLEAN — no prohibited claims detected